On your computer Local
Choose this for files and apps already on your computer.
Keep the computer awake and the agent running.
One useful task at a time
Start with a job you wish someone would take off your plate. Explain the result, give it the information, and check the work.
Follow along, copy a prompt, and try it with sample information. Replace anything in [brackets].
Asynchronous means you start a job now and come back later to review it. The AI may pause when it needs you.
Good fits include research and comparisons, summaries, organizing information, drafts, and small prototypes. These produce something you can check before using it: a supplier comparison, a service FAQ from public information, or a draft reply list.
AI can prepare options and drafts. Keep spending, consequential messages, publishing, deletion, and security or access changes behind explicit approval. You can approve specific, bounded actions; some sensitive steps require you to act yourself. Medical and financial decisions need human judgment and qualified review.
Using [approved information], prepare [specific result] for my review. Work only within [scope, time limit, and spending limit]. Include sources and unresolved questions. Do not send, publish, delete, or change access. Stop and ask if information is missing or a decision would change the plan. Finish with the result and what you checked.
Before stepping away, confirm background execution is available. Any steps using your own computer still need it awake and connected.
A long-running job needs a clear finish line. Work with AI to write the brief before asking it to build.
“Build me a virtual classroom” leaves too much to guess. Who will use it? What should happen when they click? What should it look like? An interview turns a rough idea into decisions the AI can follow while you are away.
Help me design a virtual classroom for a fictional three-year-old who is beginning to read. Interview me first. Do not build anything yet. Ask a few plain-language questions at a time about: - Learning goals, child and parent users, and activities. - Features, how they should work, parent progress tracking, and rewards. - Colors, visual style, devices, and ease of use. - Privacy, budget, time, other constraints, and what a successful first version must do. Ask follow-up questions when choices are unclear. Do not invent my preferences. Then turn my answers into a detailed, self-contained build prompt for a coding agent. Include the agreed behavior, design, limits, and checks for success. Separate the first version from future ideas and flag unresolved choices. Let me review it before implementation.
You do not need to know the technical terms. Explain the experience you want, then check that the brief describes it accurately.
Save details you will use again so you do not repeat yourself.
For a returning job, a Project can hold related chats, files, and instructions. Add the Seestar script brief and approved notes, then continue the work there.
I need a script to check my mounted Seestar S50 storage and external storage for newly captured images. Find the JPEG with the most stacked frames and copy it to a local folder. First confirm the storage locations and destination; do not invent paths. Ask what counts as new and whether I want one JPEG overall or one per observation. If frame counts are unclear, ask rather than guess. Copy selected images without changing or deleting the originals. Set up an ongoing Project to track the brief, decisions, script, and work completed. Help me version the script with GitHub. Confirm the repository before pushing code, and do not upload images unless I approve.
A Page is an editable document. ChatGPT Space holds saved work; a space groups related Pages. Ask for your approved plan as a private Page, or save a document instead.
Space is available on eligible Pro, Business, and Enterprise accounts. Create and edit on desktop or web; mobile editing is coming soon.
Check sharing first: a shared space includes its Pages, and a shared Page includes child Pages. Linked files keep their permissions; copied private details become visible to the Page’s readers.
Using the observing notes and equipment list in this project, suggest three targets for my next clear night. Ask for my location and date if they are missing. Explain why each target suits my equipment.
Create a private Space for Deep Space Field Notes, with an equipment page, an observing log, and a next-session plan. Use only the notes and files I provide. Keep source links and mark missing details as questions. Do not invent observations or change the original files. Draft the next-session plan as a Page I can edit. Ask for my location and date if they are missing. Keep everything private for my review; do not invite anyone or schedule updates.
# Working rules - Explain changes in plain language. - Label sample data and uncertain claims. - Never invent customer feedback or test results. - Test changed behavior and report what was not checked. - Finish with what changed, the evidence, and the next step.
A plugin connects AI to an app you use. Each connection gives it specific tools and permissions.
One project can use several connections. For Deep Space, each has a different job:
Help me plan the next step for Deep Space using the connections I approve. GitHub: inspect the Seestar script and its version history in [repository]. AWS Core: inspect the approved hosting setup and prepare a deployment plan. Fourthwall: check the available tools, shop permissions, and product catalog. Draft options for offering selected images for sale, if supported. Confirm the accounts, repository, AWS environment, and shop first. Say which connections are available or missing. Keep the existing Deep Space gallery and domain unchanged. Read and prepare drafts only. Ask before deploying, uploading images, publishing listings, changing prices or domain records, or spending. Report what you verified and what remains proposed.
Connecting an account is different from approving a change. Start with reading and a plan; review before anything is deployed, published, or offered for sale. Available plugins and actions vary by account.
Choose the account and access this project needs. A connection can be installed but still need sign-in or permission.
Confirm the selected repository, hosting account, or shop. Reading information and changing it are separate permissions. Do not use a write action just to test the connection.
A Skill saves a method you want AI to follow again.
Save a Weekly Update skill. Ask which approved notes to use. Summarize completed work, blockers, decisions, and next actions. Include sources. Flag missing facts; do not invent progress, owners, or dates. Keep it as a draft. Do not send, change shared files, or schedule anything. Test once with sample notes. If saved skills are unavailable, give me a reusable prompt.
A Skill remembers how; a schedule decides when. Saving a Skill does not grant access or start future work.
Help me create two reusable skills: Good Morning and Good Evening. Ask which calendar, email, chat, and document connections to check, and where to save my daily note. Good Morning: check those sources and give me three priorities, today's meetings, and blockers. Good Evening: compare the morning plan with actual progress. Save completed work, open commitments, and tomorrow's first step. Include source links and say when a connection could not be checked. Do not send messages or change my calendar. Test both skills once.
In ChatGPT, choose @skill-creator where available. In Codex CLI or the IDE, use $skill-creator. Claude has its own skill settings and Claude Code uses SKILL.md files.
A Site turns your work into a website or interactive tool people can open and use.
Deep Space turns selected telescope images and capture details into a gallery. Its existing project keeps image preparation, review, and publication separate. Build on that work rather than start over. See the example project.
Help me build on my Deep Space gallery to share selected Seestar images and their capture details. First inspect [confirmed GitHub repository] to understand the existing script and gallery. Confirm the image folders, selected images, available metadata, preview location, and publication destination. Do not invent paths or missing facts. Use approved image copies. Preserve the originals, existing gallery and domain, and private working files. Show verified capture details beside each image; mark unknown details clearly. Prepare a reviewable preview for phone and desktop. Test image loading, navigation, and captions. Explain the changes and publication plan. Wait for my approval before uploading images, deploying, publishing, changing the domain, or adding sales links.
Open the preview and try it. Check the images, captions, phone layout, and who can access it. A working preview is a review point; publishing is a separate decision.
Where enabled, a workspace-private Site can use supported plugins to show each signed-in visitor data from their own approved accounts. Sharing the Site does not share your account connections. Sharing, plugin access, and scheduled updates have separate settings.
@Sites Create a short, spooky browser escape game called Before the Morning Bell. Start with one room, one hidden key, and one locked exit. Add a clear goal, readable controls for mouse and touch, and an ending when I escape. Include an optional sound toggle; keep sound off until I turn it on. Save a first version for me to review before publishing. Include brief instructions for trying it.
A scheduled task repeats a job at a saved time. Test the job once before scheduling it.
A daily fantasy review can turn fresh league information into decisions for you to make. What matters changes during the week: completed games, waivers, injuries, or the next lineup deadline.
Use my approved fantasy connection to review both of my Sleeper teams. Confirm the teams and current season/week. Keep each league’s rules and recommendations separate. Check recent results, upcoming matchups, rosters, injuries, available waiver options, and lineup suggestions. Focus on today’s useful decisions and verified deadlines. Include current sources, an as-of time, and missing data. Do not invent results or availability. Recommendations only: do not change lineups, add or drop players, submit claims or bids, or make trades. Run once for my review. Then help me schedule this every day at [time] in [time zone], delivered privately to [destination]. Confirm both teams, source access, and the saved task. If scheduling is unavailable, give me the prompt to run manually.
Check the actual saved schedule, time zone, destination, and first result. A written promise is not a created task. Writing a cadence on a Page does not schedule it. The task must be able to reach both teams’ data when it runs.
Check these dumpster rental companies in [city]: [official website URLs]. Compare [dumpster size], rental days, included weight, base price, and extra fees. Link each source and date the result. Mark unpublished prices “quote required.” Do not contact the companies. Run this once now and save the comparison in [destination]. After I check it, schedule the same task for Mondays at 8 AM in [timezone]. Compare with the last successful result and report changes or failed checks. Confirm the task exists and where I can see its results.
Give one outcome, a checkpoint, and a clear finish line.
Use [approved brief and files] to prepare an event plan, invitation draft, reply queue, and weekly update. Reuse the same verified facts; label unknowns. Show the plan and wait for approval. Then save each finished draft. Pause when a missing decision blocks progress. Do not send, book, buy, publish, or change sharing. Finish with the four drafts, sources, checks, and unresolved questions. Tell me where to reopen this task.
At a checkpoint, ask: “Compare this with my brief. What passed, failed, or still needs review?” Correct the scope, then resume the same task.
Cloud work runs on a remote computer. Eligible new Work tasks can continue across devices with sync enabled. Check that the task can reach its files and tools; cloud access does not include your local files or sign-ins automatically.
Steps using your computer need it awake, online, and connected. Any job can pause for missing inputs, permission, or limits.
Example · BullyBearAI An overnight Codex session worked through epic #1090 and produced eight draft pull requests—proposed code changes ready for review.
The run reported 11h 25m of session elapsed time and a separate 2h 46m goal timer. It did not deploy to AWS. The Claude edition adapts the workflow; the original run used Codex.
Codex Cloud currently does not support browser or computer use. Work and your dot’s cloud computer have different capabilities. Save important output and commit code changes; saved task state is not a replacement for source control. Cloud task guide · Environment setup
Start in Codex with the project and required access ready. For a job that should run while your laptop is closed, use a prepared Codex Cloud environment. You can check its progress from another device. Local or Remote work still needs its host computer available.
Use /goal where supported, including the desktop app, interactive Codex CLI, or IDE extension. In Work on the web, paste the brief directly. Give either tool the same clear outcome, boundaries, and completion checks.
In Codex, work through the AFK-ready features under [repository and epic]. Refresh the issues and project rules first; skip blocked or human-dependent work. Use up to two subagents on independent tasks, with separate branches. For each feature, implement it, run the relevant tests, update documentation, and open a draft pull request. Save progress after each feature. Stop after [time limit], at [usage limit], or when no ready work remains. Report any limit you cannot measure. Do not merge or deploy. Finish with PR links, test evidence, dependencies, and what needs a person.
A second chat gets a clean look at the work. Include the epic, pull requests, and the original completion checks. If deployment is part of the review, name the exact test environment.
Review BullyBearAI epic #1090 and PRs #1128, #1129, #1130, #1134, #1135, #1136, #1139, and #1140. Fetch current changes and dependencies; do not rely only on the author's summary. Check the code and rerun tests. Fix problems on the affected PR branch and push the updates. Recheck after every fix. Deploy and test sequentially only in [authorized AWS test account, region, and environment]. Stop if that target is unspecified. Record the deployed commit and results for each PR. Do not merge or touch production, broker access, or trading controls. Finish with the review order and remaining blockers.
Before you leave: check that the agent can reach the files and tools, knows when to stop, and has somewhere to save progress. Long jobs can still stop for limits, questions, or missing access.
A dot is an ongoing assistant that can coordinate work and bring back decisions.
Create your dot on desktop or web, then use the supporting mobile app update. Review its plugin permissions and use Custom Rules to set approval boundaries.
Help me prepare [product] for launch on [date]. First ask about the customer, readiness, and sources you may use. Draft a launch checklist, five customer answers, and three blockers. Link evidence and flag missing facts. After I approve the plan, keep these drafts current through launch day and tell me what changed. Do not send, publish, spend, or change code. Report access problems or limits. Stop on launch day unless I extend the work.
Its cloud work can continue between conversations. Local-computer steps still need that computer available. Background research is read-only; actions follow your permissions and safety rules.
Rolling out to eligible Pro and Business Premium accounts. Enterprise, Edu, and Healthcare need admin-enabled beta access. If unavailable, try a one-off Work task or Chat. Dot availability and setup
Before you call it done
Open the result. Try it. Check facts and missing work. Save the instructions that helped.
Start with the AI you already use. Chat handles questions and drafts. Work handles jobs with files and tools. Codex is a dedicated workspace for building and testing software. A dot can coordinate ongoing work; it still needs the right access and permissions.
Features vary by plan, app, region, and workspace settings. If a feature is unavailable, practice the same brief in Chat. You can ask for a website without learning to read code.
GitHub stores software projects. AWS provides cloud computing services. Neither is required for these everyday exercises.
DevDay 2026. Checked September 30. Beta, preview, and coming-soon features are labeled; account and rollout limits apply.
Sources: September 29 DevDay recap, DevDay product guide, and the feature documentation linked above. Current docs supply the practical restrictions where announcement summaries are broader.
A practical guide. One useful task at a time.
Turn a postponed job into a clear brief, let AI do useful work, and check the result. Start with a document, an email draft, or an event plan.
Follow the lessons below. Copy a prompt, replace the bracketed details, and try one task. Use sample information while learning. No coding or app connections are required to begin.
Start with the kind of task you want done. You do not need every tool to begin.
Choose this for files and apps already on your computer.
Keep the computer awake and the agent running.
Choose this for work that can use uploaded files and connected online tools.
Your laptop can be off if no step needs a local tool.
How to tell: in Claude Desktop’s Code tab, check the Local / Cloud environment selector. A normal terminal session runs on that computer. Remote Control still depends on its host computer. For other tasks, check whether the tools use local folders or desktop apps.
A cloud job can still pause for missing information or permission. Labels and features vary by account and app version.
Describe what should exist when the work is finished, what the AI may use, and where it must stop.
Example · A community open house Ask for a plan you can review, with unknown costs and decisions clearly marked.
A good brief names the outcome, permitted inputs, constraints, approval boundary, finish line, and supporting evidence.
Help me prepare a small community open house. Outcome: A one-page event plan and an invitation draft for my review. Inputs: This fictional example only: October 15, 5:30–7:00 p.m. Eastern; 30 expected guests; a confirmed room; a $600 planning limit. Refreshment prices and accessibility details are unknown. Constraints: Keep it simple. Label unknown costs and suggested deadlines. Do not invent facts. Approval boundary: Do not send, book, buy, publish, or invite anyone. Finish line: The plan includes a schedule, a checklist, and the decisions I still need to make. Evidence: Show which details came from my brief and which are suggestions. First ask up to three questions that would materially improve the plan. Then show a short plan and wait for my approval before doing the work.
“Build me an LMS” (a learning management system) leaves a lot to guess. Spend time describing what you want before asking the agent to build it.
Example · Izzy’s Virtual Classroom A 3D homeschool world: finishing a real book adds it to a virtual bookshelf, unlocks a reward, and gives parents a record of learning.
Help me write a build prompt for my daughter's learning management system: a playful 3D homeschool classroom with reading, math, science, and a useful parent view. Interview me first, 3–5 questions at a time. Do not start building yet. Ask follow-up questions when my answers are vague. Help me define: - The experience: who uses it, what they do, and how it should feel. - The deliverable: the first complete learning activity, what must work, and how we will test it. - The guardrails: privacy, child safety, budget, and what is out of scope. - Authorized actions: which files and tools you may use, what you may change or install, and whether you may spend money or publish. Then turn my answers into one self-contained implementation prompt. Separate the first version from future ideas, list unresolved assumptions, and let me review the prompt before implementation.
This is the implementation prompt for after the interview. Notice the specific learning activity, parent tools, privacy rules, permissions, and completion checks. Adapt the personal details and permissions before using it.
I want to build a virual classroom website for my 3 year old daughter. we've started homeschool. Build “Izzy’s Virtual Classroom” — A 3D Homeschool Learning World You are being given this project as a one-shot autonomous engineering task. Your job is not merely to scaffold a repository, create a mockup, write a design document, or produce disconnected prototypes. Build the most complete, polished, runnable version of the product that you can within the available environment. Act as the product designer, senior full-stack engineer, game developer, curriculum-systems architect, UX designer, and QA engineer. Do not wait for clarification unless something is literally impossible to infer. Make strong, sensible product decisions yourself. The primary evaluation criteria are: How much of the vision you successfully implement. How polished and cohesive the finished experience feels. Whether the core gameplay loop actually works. Whether the application feels like a real product rather than a developer demo. Code quality, architecture, maintainability, performance, and tests. How intelligently you handle parts of the vision that cannot reasonably be completed in one pass. PRODUCT VISION Build a personalized 3D homeschool world initially for a child named Izzy. Izzy turns four on January 1. Despite her age, she is already reading independently at approximately a third-grade level. Her parents want to challenge her intellectually without forcing every subject into a conventional age-grade box. The system should recognize that a child may simultaneously have very different ability levels in: reading mathematics science writing reasoning vocabulary life skills creativity The application should therefore adapt by domain, not assign one global “grade level.” The experience should feel like a genuine video game that happens to provide a rigorous education. It should NOT feel like: a homeschool spreadsheet with cartoons a generic LMS a collection of worksheets a dashboard with “gamification” a preschool app full of giant primary-colored buttons a thin Three.js tech demo The central fantasy is: Izzy has her own living 3D school. She walks through it, learns inside it, reads real books, completes lessons and real-world homeschool activities, earns things through genuine progress, watches her school physically grow as she learns, and can proudly show her family everything she has accomplished. The emotional target is this moment: Izzy runs to the computer because she wants to put the new book she finished onto her virtual bookshelf, sees that she has now read 100 books, receives a new pet dragon, and then excitedly gives Grandma a tour of everything she learned that month. Build toward that feeling. DESIGN DIRECTION The visual direction should combine: A personalized dream school with: a beautiful, playful, believable elementary-learning environment. Think: warm cozy tactile inviting magical without being excessively fantasy-themed expressive lighting high-quality materials playful environmental storytelling little visual surprises rooms that look like places a child would actually want to explore Do NOT imitate copyrighted worlds or characters. Avoid making it feel like Hogwarts, Minecraft, Roblox, Fortnite, Duolingo, etc. It should have its own identity. The world can gradually become more fantastical as Izzy progresses. For example, an initially grounded science room might eventually contain: a working planetarium a giant aquarium a dinosaur exhibit a glowing crystal collection a telescope observatory a greenhouse Learning should literally transform the environment. PRIMARY USER EXPERIENCES There are three primary modes. 1. CHILD MODE This is Izzy’s experience. It should be a true explorable 3D game environment. She should be able to: walk around explore rooms interact with objects approach teachers inspect her bookshelf begin lessons see accomplishments unlock objects interact with rewards see her school physically change over time Controls should be simple enough for a young child. Support keyboard + mouse initially. Structure controls so controller support can be added cleanly. Avoid interaction patterns requiring precision mouse use. Large interaction ranges and obvious affordances are preferable. 2. PARENT MODE The parent-facing experience is primarily for Izzy’s mother. This needs to be extremely useful and should not be treated as an afterthought. Parents need a serious homeschool recordkeeping and curriculum tool underneath the game. The parent should be able to record things like: “Izzy read two chapters of Charlotte’s Web this morning. She summarized what happened without help and correctly explained why Wilbur was scared. Later we baked bread and she measured 2 cups of flour and 1 tablespoon of yeast.” The system should interpret this into proposed educational evidence such as: Reading Reading comprehension Oral narration Vocabulary Measurement Mathematics Science Life skills The parent should then be able to review/edit/save it. For this implementation, if a live AI model is unavailable, implement a believable local rules/heuristics system and a clean abstraction for replacing it with an AI service later. Do not make the whole feature dependent on having an API key. Parent workflow should emphasize: type one thing → system does the organization rather than forcing the parent to complete a giant form. 3. FAMILY SHOWCASE Create the beginnings of a mode where Izzy can show someone what she has accomplished. This may be called something like: My Learning Museum My School Izzy’s World Showcase A family member should eventually be able to see things such as: books completed projects science discoveries artwork rooms unlocked milestones photos favorite accomplishments For this first implementation, build a useful local showcase even if remote sharing is not yet implemented. CORE GAME LOOP The most important flow is: Izzy enters her school. She explores the 3D classroom. She enters or approaches the library area. She interacts with a virtual teacher. She records or completes progress on a real book. That book visibly appears on her physical 3D bookshelf. Her reading collection/progress changes. She earns a meaningful reward. Her world changes in some visible way. Parent mode shows the activity and learning evidence. This loop must work. If you have to choose between implementing 20 half-working systems and making this loop excellent, prioritize this loop. WORLD STRUCTURE The school should be designed to expand over time. Initial or future areas include: Central Classroom Library Math Room Science Lab Art Studio Garden / Greenhouse Observatory Learning Museum Reading nook Trophy / accomplishment area Not all rooms need to be fully implemented immediately. However, build the architecture so rooms are modular and unlockable. A good first playable environment may contain: main classroom library zone science corner locked doors or visible future spaces The environment should communicate: “There is much more to discover.” WORLD PROGRESSION Learning should alter the world. Examples: Read 10 books: bookshelf expands Read 25 books: reading nook unlocks Read 50 books: larger library opens Read 100 books: special celebration + pet dragon Study plants: greenhouse gains plants Complete a space unit: observatory unlocks objects Study dinosaurs: museum gains a fossil or dinosaur display Do a nature walk: collected items can populate a nature exhibit Complete artwork: artwork can appear on classroom walls The school itself should become a persistent visual record of learning. Implement at least several examples of this system. READING SYSTEM Reading is especially important. Every real book Izzy reads should be capable of becoming a persistent object in her virtual library. Each book record should support fields such as: title author cover image if available date started date completed pages or chapters child rating parent notes comprehension notes favorite part optional drawing/work sample independent / assisted reading tags reading difficulty metadata if known The child-facing experience should NOT obsess over grade labels. Izzy should largely be free to read books she loves. Internally, parents may record or review difficulty and ability evidence. The 3D shelf should visually fill as books are completed. Make this visually satisfying. If external book APIs are unavailable, include seeded sample books and build an API abstraction for future metadata lookup. CURRICULUM The application should actually teach. Primary academic focus: Reading Mathematics Science Secondary domains may include: writing art nature life skills problem solving vocabulary Underneath the playful experience, curriculum should map to established educational frameworks where appropriate. Examples: Mathematics Use concepts compatible with Common Core-style progression. Science Use concepts compatible with NGSS-style progression. Literacy Use evidence-based literacy concepts and age-independent reading comprehension skills where appropriate. Do not expose standards codes to the child. Example: PARENT VIEW: Number & Operations Demonstrates addition within 20 independently. CHILD VIEW: Rocket Rescue Complete! 🚀 Create a curriculum data structure supporting: domains strands skills prerequisites mastery state evidence difficulty attempts parent observations lesson relationships Seed enough curriculum content to demonstrate the system meaningfully. ADAPTIVE LEARNING Do NOT give Izzy one global grade level. Maintain independent capability estimates per subject or skill domain. For example: Reading comprehension: advanced Phonics: mastered foundational concepts Addition: developing Measurement: emerging Scientific observation: strong The system should attempt to present work just beyond demonstrated mastery. A child who struggles should receive: hints visual explanations manipulatives simpler intermediate problems alternate explanation another attempt Avoid punitive failure. Do not use: harsh red Xs lost points humiliation “you failed” punishment for mistakes Use language like: “Almost. Let’s try another way.” However: Do NOT fake mastery. The parent system should accurately distinguish: introduced developing proficient mastered Create the foundations for a lightweight adaptive engine. A deterministic implementation is acceptable for this first version. FORMATIVE ASSESSMENT Assessment should often be disguised as play. Do not constantly announce: TEST TIME Examples: A reading teacher asks Izzy what happened in the story. A math robot needs help distributing 14 moon rocks. A science teacher asks which object will sink. Use results as evidence of skill. Build at least one functioning example. The system should preserve evidence that can support statements like: “Izzy independently answered inference questions about this text.” rather than merely saying: “Izzy was exposed to reading.” VIRTUAL TEACHERS Include recurring subject teachers/NPCs. Suggested starting cast: Professor Hoot Reading and storytelling. Personality: warm, curious, enthusiastic about books. Digit Math. Could be a small friendly robot. Personality: playful, precise, loves puzzles and counting. Nova Science. Could be an adventurous animal, creature, or young scientist. Personality: curious, experimental, constantly asking “why?” You may improve these names/designs if you have a stronger cohesive art direction. Teachers should: have recognizable locations greet the child provide lesson interactions give encouragement present learning objectives through stories/tasks adapt difficulty remember basic progress Do not build unrestricted AI chat for the child. Teacher interactions should initially be: bounded educational lesson-aware predictable parent-visible Design the architecture so conversational AI can later augment structured interactions. VOICE-READY ARCHITECTURE Eventually, voice may become a major interface. Potential future interaction: Teacher: “What do you think will happen if we put this in water?” Izzy: “I think it sinks because it’s heavy.” The system could assess reasoning and continue. Do not make voice recognition mandatory for this version. But architect teacher interactions so text input can later be replaced or supplemented by: speech-to-text text-to-speech pronunciation analysis oral narration conversational tutoring If browser-native speech support can be safely and cleanly implemented as an optional enhancement, you may add it. Always provide a non-voice fallback. REWARDS Rewards should celebrate genuine learning. Examples: pet companions classroom furniture plants art outfits trophies bookshelf expansions room expansions observatory equipment science objects museum exhibits environmental effects Avoid manipulative mobile-game mechanics. NO: loot boxes monetization paid currencies artificial scarcity punishment for missing a day anxiety-producing streak loss endless notification mechanics Streaks may be displayed positively if useful but should never punish absence. Prefer: Knowledge → Collections → World Growth over generic XP grinding. PET SYSTEM Implement at least the foundation of collectible companions. A pet dragon should exist as a high-profile reward, even if only as a milestone demonstration. Example: 100 books read → special dragon companion unlocked. For development/demo purposes, include a debug/demo way to preview milestone rewards without manually entering 100 books. Do not undermine the actual progression system—this should simply aid testing. REAL-WORLD HOMESCHOOL ACTIVITIES Parents should be able to record activities outside the computer. Examples: baking nature walks gardening museum trips science experiments handwriting painting building something counting money helping cook observing wildlife These should contribute to the learning portfolio. Activities should support: date narrative subjects skills photos/work samples parent observations child reflection standards alignment demonstrated mastery/evidence duration if desired The UI should make these easy to enter. PARENT DASHBOARD Create a polished parent experience. At minimum include: TODAY recent activities lessons books progress suggested follow-up PROGRESS Break down domains such as: Reading Math Science Writing Creativity Life Skills CURRICULUM Show: mastered proficient developing introduced suggested next concepts PORTFOLIO Show: photos projects work samples book records assessments observations REPORTS Create at least one usable report view. Eventually support: weekly monthly semester annual grandparent-friendly showcase For this version, generate a polished learning summary using locally stored information. NATURAL LANGUAGE ACTIVITY ENTRY This is a major feature. Parent enters: “Izzy and I baked muffins. She counted 12 cups, measured the flour herself, read several steps of the recipe, and asked why the muffins get bigger in the oven.” The system should infer or propose: Math Counting Measurement Reading Sequencing Science Chemical/physical change curiosity Life skills The parent reviews the suggestions before saving. If no LLM is configured, implement a local parser/classifier good enough to demonstrate the workflow. Architect it behind a service interface such as: ActivityInterpretationService so a real AI provider can later replace it cleanly. MULTI-CHILD ARCHITECTURE Although Izzy is the initial child, this must be built as a multi-child household system. Her younger sister Georgia should eventually have her own profile. Architecture: Household → Parent accounts → Child profiles → Independent curriculum state → Independent worlds → Independent books → Independent portfolios → Independent rewards The children may eventually share family spaces, but progress must remain separable. Seed Izzy as the initial profile. Georgia can exist as an optional/inactive future profile. AVATAR SYSTEM Eventually, parents should be able to upload a photo selected by the child and use it as inspiration for a game avatar. For this version: Build avatar customization architecture. Provide a starter Izzy avatar. Allow customization such as hair, clothing, skin tone, accessories, etc. Provide a parent-controlled photo upload area if practical. Do NOT implement biometric identification or face recognition. Do NOT automatically upload children's photos to third-party services. Keep any future image-to-avatar pipeline clearly separated behind a safe service boundary. If image transformation is unavailable, do not fake it. Instead make the UX clearly show how it would fit later. VISUAL QUALITY This application should look intentional. Do not ship a gray-box scene with cubes and default materials and call it complete. Spend time on: lighting materials environmental composition UI typography camera feel interaction feedback transitions animations sound hooks environmental storytelling Use a cohesive art direction. Stylized realism is preferable to attempting photorealism. Think polished indie game, not enterprise dashboard. 3D TECHNICAL DIRECTION Prefer a web-first architecture. Recommended stack: React TypeScript Vite or appropriate modern equivalent Three.js React Three Fiber Drei where helpful Zustand or similarly lightweight state management if useful Use physics only where it adds value. Do not add unnecessary complexity. Target: desktop browser first keyboard/mouse architecture compatible with later controller support eventual tablet compatibility Keep rendering performance reasonable. Use: instancing where appropriate optimized meshes compressed assets where practical sensible shadow settings lazy loading LOD only if necessary BLENDER Use Blender as an asset creation pipeline if it is available and useful. You are authorized to: inspect whether Blender is installed use Blender scripting generate models procedurally create/modularize classroom assets export GLB/glTF assets build low-poly/stylized environmental pieces If installing Blender is permitted and reasonable in the execution environment, you may install it. Do NOT block the project on Blender. If Blender is unavailable: create high-quality procedural geometry or temporary assets in Three.js preserve an asset pipeline that allows Blender models to replace them later If you use Blender, save source .blend files or generation scripts in a clear assets/source directory. SOUND Add tasteful game audio architecture. Useful examples: footsteps UI feedback page/book sound unlock sound gentle classroom ambience teacher greeting cues Do not use copyrighted music. Do not make audio annoying or constant. Provide volume/mute controls. If actual sound assets are unavailable, create the architecture and use legally safe generated/simple audio where possible. DATA ARCHITECTURE The project should use a real persistent data model. For a local-first first version, acceptable approaches include: SQLite IndexedDB local backend database another sensible embedded/local datastore Prefer a solution that can migrate cleanly to production. Potential entities: Household Parent Child Avatar Book ReadingSession Activity Evidence Subject Skill CurriculumStandard Lesson LessonAttempt Mastery Reward UnlockedItem Room PortfolioItem Report TeacherInteraction Do not make everything a single giant JSON blob. Use sensible relationships and types. PRIVACY AND CHILD SAFETY This application stores information about young children. Design accordingly. Requirements: parent is the administrative authority avoid unnecessary personal information no ad tech no trackers no social network no public child profiles no unrestricted web browsing for the child no child-facing external links no manipulative engagement mechanics parent can inspect educational interactions photo uploads remain controlled by the parent do not use face recognition do not silently send child data to third parties structure external AI integrations as explicit optional services A local-first architecture is strongly preferred for this prototype. DEMO CONTENT Seed enough data for the product to feel alive immediately. Include examples such as: several completed books a partially completed book several math skills several science skills a baking homeschool activity a nature activity a reading assessment progress toward an unlock at least one unlocked cosmetic/world reward Do not make Izzy's fictional educational history look authoritative. Clearly treat seeded information as demo/sample data. Provide an easy way to reset demo state. VERTICAL SLICE The highest-priority deliverable is a complete vertical slice. It must include this sequence: CHILD Start application. Select/enter Izzy’s world. Spawn inside a polished 3D classroom. Walk around. See a recognizable library area. Approach Professor Hoot or equivalent reading teacher. Interact. Choose/add/finish a book. Answer at least one small comprehension/activity interaction. Complete the interaction. Watch the book appear physically on the bookshelf. See reading progress change. Receive a reward or unlock. See the environment react. PARENT Enter parent mode. See the reading activity. See evidence from the interaction. Enter a natural-language homeschool activity. Review automatically proposed subject/skill classifications. Save it. See progress/portfolio update. Open a learning report summarizing recent progress. This entire flow should be functional. GAME FEEL Pay attention to feel. Implement small details like: smooth movement camera damping hover/interact cues object highlighting satisfying unlock animations environmental animations NPC idle motion bookshelf changes animated doors if appropriate particles used sparingly polished menus child-friendly readable typography obvious interact prompts Avoid overloading the screen with HUD elements. The world itself should communicate progress whenever possible. CHILD UX Remember: Izzy is almost four. Even though she reads far beyond her age, motor/UI expectations should remain developmentally appropriate. That means: large interaction targets minimal nested menus visual cues few simultaneous choices forgiving controls little required typing optional reading aloud later clear back/home navigation no complex settings exposed in child mode Advanced reading ability does not mean adult UI dexterity. PARENT UX Parent mode may be much denser. Optimize it for: speed clarity evidence planning records Parents should be able to understand: “What does she know?” “What is she working on?” “What did we do?” “What should we try next?” “What evidence do we have?” within a few minutes. ENGINEERING QUALITY Use: strict TypeScript clear components/modules sensible separation between game, learning domain, and persistence automated tests formatting/linting error handling accessible non-3D parent UI deterministic seeded data clear environment configuration Do not scatter magic constants throughout the code. Create appropriate abstractions for: persistence curriculum lesson engine adaptive progression rewards AI activity interpretation asset loading teacher interactions Avoid premature microservices. This should be easy to run locally. TESTING Test critical logic. At minimum test areas such as: mastery/progression transitions reward unlocking book completion activity classification multi-child data separation persistence curriculum recommendations Where practical, add browser/end-to-end tests for the main vertical slice. Actually run tests. Fix failures. QUALITY ASSURANCE Before considering the task complete: run the app inspect it visually click through major flows verify 3D controls test book completion verify bookshelf update verify reward unlock test parent entry test report generation test persistence after reload check console for errors check responsive parent UI run automated tests run lint/type checking fix obvious visual defects Do not claim a feature works because code exists. Verify it. PROJECT DOCUMENTATION Include a useful README. It should contain: what the product is architecture setup run commands test commands major features implemented major directories data model overview curriculum architecture 3D asset pipeline Blender workflow if used how to replace local activity interpretation with a real LLM future voice architecture known limitations recommended next development priorities Also create a concise architecture/design document if helpful. AUTONOMY This is deliberately a one-shot test. Do not stop after scaffolding to ask what to do next. Do not ask me to pick: libraries colors schemas folder names state management database implementation art direction details component architecture Decide. If one approach fails, try another. If a dependency is unavailable, create a fallback. If you cannot implement the entire vision, prioritize a cohesive, impressive vertical slice and leave the repository substantially better than a prototype. Use the available tools and execution environment aggressively. Inspect your own work as you go. PRIORITY ORDER If time/capacity forces tradeoffs, prioritize in this order: Runnable application. High-quality 3D classroom. Good player controls/interactions. Library and physical bookshelf. Book completion loop. Virtual reading teacher. Reward/world-growth system. Parent activity logging. Natural-language interpretation. Evidence/portfolio model. Curriculum and mastery model. Parent progress dashboard. Learning report. Math/science lesson examples. Additional rooms. Avatar improvements. Audio polish. Additional world content. Quality of the core experience is more valuable than breadth. SUCCESS CRITERIA When you finish, I should be able to clone/open the project and reasonably say: “This is the beginning of a real 3D homeschool game, not a mockup.” Izzy should be able to: walk through her classroom, interact with a teacher, finish a book activity, see that book join her growing physical library, and receive something exciting for learning. Her mother should be able to: record homeschool experiences naturally, preserve detailed evidence, see meaningful academic progress, and produce a useful learning summary. And the architecture should make it believable that this could grow for years with Izzy and later Georgia. FINAL EXECUTION INSTRUCTIONS Start by inspecting the current environment/repository. If the repository is empty, initialize it appropriately. Then implement. Do not spend your response describing what you intend to build instead of building it. Create files, install dependencies where appropriate, run commands, generate assets, test the system, inspect results, and iterate. Use the entire available work session productively. At the end, provide a concise handoff containing: What you built. What works end-to-end. Architecture and important implementation decisions. Exact commands to run it. Tests/checks performed and their results. Anything that remains incomplete. The highest-value next improvements. The code and working application are the deliverable. Begin.
Clear decisions matter more than word count. A complex project deserves a detailed conversation. Save the prompt that works as a playbook you can adapt next time.
Give ongoing work one home for its context, then save the useful result where you can return to it.
A Project keeps related conversations, source files, and instructions together in Claude. Start each distinct job in its own chat inside that Project.
Example · Community Open House Keep the event brief, sample messages, approved plan, and follow-up drafts together so you do not explain the event again.
Use this Project for our community open house. Keep the event brief, sample messages, and approved plan together. Project instructions: Use only the sources I provide. Mark missing facts as questions. Keep invitations as drafts. Ask before sending, spending, publishing, or changing sharing. Include source names or links when you report a fact.
Use Project instructions for tone, sources, and actions requiring approval. Software projects can store standing instructions in CLAUDE.md, a text file read by Claude Code.
Using the observing notes and equipment list in this project, suggest three targets for my next clear night. Ask for my location and date if they are missing. Explain why each target suits my equipment.
# Working rules - Explain changes in plain language. - Label sample data and uncertain claims. - Never invent customer feedback or test results. - Test changed behavior and report what was not checked. - Finish with what changed, the evidence, and the next step.
Use information you have permission to share. Add app access only when it helps with the task.
A plugin can add tools, app connections, or reusable skills. A connector gives access to an app. You decide which account and permissions to allow.
| Connection | What it makes possible |
|---|---|
| Gmail | Find customer emails and prepare replies. |
| Google Drive | Read plans and documents without attaching each one. |
No connection needed to practice: paste fictional notes or attach a safe sample document. Connect an account only when the real job needs it and your organization allows it.
Prepare a reply queue from these fictional event messages. Show the question, facts needed, and a draft response only when the information supports one. Alex: What time should I arrive to help set up? Jordan: The room is confirmed for October 15, 5:30–7:00 p.m. Casey: Is the entrance step-free? Automated notice: There are 22 registrations so far. Setup access and accessibility details are not confirmed. Flag them instead of guessing. Account for all four messages. Do not connect to an inbox, send replies, or change anything.
Start with one app such as Gmail or Google Drive. Check its account and permissions before approval. Reading a message and sending one are different actions.
Find recent customer questions in my connected Gmail and the product plan in Google Drive. Summarize the three biggest gaps between what customers ask for and what we plan to build. Link the sources. Only read; do not send or edit anything.
GitHub stores software projects and proposed code changes. AWS provides cloud computing services. Neither is a prerequisite for this guide’s everyday exercises.
| Optional tool | What it helps with |
|---|---|
| GitHub | Add repository files as context. For issues and pull requests, use Claude Code with authenticated GitHub tools. |
| AWS tools | Use an AWS MCP connection or authenticated AWS CLI in Claude Code to inspect cloud services. |
A Skill is reusable instructions for a kind of work. Start with a method you have already tested.
Example · Weekly Update Turn your notes into the same useful report each week, with sources and open questions.
Help me save a reusable Weekly Update skill. Each time I run it, ask which notes or approved sources to use. Produce a short update with completed work, blockers, decisions needed, and next actions. Link each factual claim to a source. Label suggested actions and missing information. Do not invent progress, owners, dates, or approvals. Keep the result as a draft for me. Do not send it, edit shared files, or schedule future runs. Test the skill once using sample notes. If saved skills are unavailable, give me the instructions to reuse as a prompt.
Try it once: where saved Skills are available, ask Claude to save these instructions and then run the Skill on sample notes. Otherwise, keep the prompt in your Project or notes and reuse it manually.
A Skill remembers how. A scheduled task decides when. A plugin supplies tools or access. Saving instructions does not schedule work or grant permission.
Help me create two reusable skills: Good Morning and Good Evening. Ask which calendar, email, chat, and document connections to check, and where to save my daily note. Good Morning: check those sources and give me three priorities, today's meetings, and blockers. Good Evening: compare the morning plan with actual progress. Save completed work, open commitments, and tomorrow's first step. Include source links and say when a connection could not be checked. Do not send messages or change my calendar. Test both skills once.
In ChatGPT, choose @skill-creator where available. In Codex CLI or the IDE, use $skill-creator. Claude has its own skill settings and Claude Code uses SKILL.md files.
Choose an output you can open, use, and check. A useful first version can be small.
A Artifact can turn an idea into something you can try, such as a journal, checklist, or small interactive tool. Claude’s Artifact tools and publishing options differ from ChatGPT Sites.
Example · Winery Journal Keep a few fictional visits, ratings, and notes in a simple tool you can test on your phone.
Create an interactive Artifact for a simple Winery Journal using fictional data. Let me review it before publishing. Use three clearly labeled fictional wineries. Include visited / want-to-visit status, a 1–5 rating, a short note, search, and a way to add or edit an entry. Make it easy to use on a phone. Show the visited count and average rating, excluding unrated entries. Explain where edits are saved and whether they persist after refreshing. Do not invent real business facts. Skip login, payments, maps, and external accounts. Do not publish or invite anyone. First show a short plan and wait for approval. Then build the preview and report what you tested. If a preview requires publishing, pause and ask.
Test the actual result. Add a sample entry, edit it, try the filters, and refresh. Check where data is saved before trusting it with anything important.
Create an interactive Artifact: a short, spooky browser escape game called Before the Morning Bell. Start with one room, one hidden key, and one locked exit. Add a clear goal, readable controls for mouse and touch, and an ending when I escape. Include an optional sound toggle; keep sound off until I turn it on. Show a playable Artifact and explain how to try it. Let me review it before publishing.
Repeat a proven job at a useful time. Availability and source permissions still apply.
A task is a piece of work you assign. A scheduled task runs later or repeats at a saved time. Browsing lets it check current information on websites.
Example · Friday event update Review the approved notes and prepare a private summary of progress, blockers, and decisions.
First run this job once using [approved notes or sources]: prepare a private weekly update with completed work, blockers, decisions needed, and next actions. Include sources and uncertainties. Do not send it to anyone else. After I review the result, help me schedule it for Fridays at 3 p.m. in [my time zone]. Confirm the sources, destination, exact schedule, and saved task before saying it is active. If scheduling is unavailable, give me the prompt to run manually.
Confirm the saved task. Check its time zone, sources, destination, and first result. A written promise is not a created task. Use cloud-accessible sources for work that should run with your laptop off.
Check these dumpster rental companies in [city]: [official website URLs]. Compare [dumpster size], rental days, included weight, base price, and extra fees. Link each source and date the result. Mark unpublished prices “quote required.” Do not contact the companies. Run this once now and save the comparison in [destination]. After I check it, schedule the same task for Mondays at 8 AM in [timezone]. Compare with the last successful result and report changes or failed checks. Confirm the task exists and where I can see its results.
Delegate an outcome across several steps, with checkpoints and a clear stopping condition.
Example · Community open house packet One job can produce an event plan, invitation draft, reply queue, and weekly update from the same approved notes.
Prepare a review packet for our community open house from [approved brief and sample files]. Include a one-page event plan, an invitation draft, a reply queue, and a short weekly update. Reuse the same verified event facts throughout. Keep unknown costs, access details, and unassigned work clearly marked. First show the plan and wait for approval. After approval, work through the packet and save each completed draft. Pause if a missing decision materially changes the result. Do not send, book, buy, publish, or change sharing. Finish when all four drafts are ready for review or a required input blocks progress. Report source files used, checks performed, and unresolved questions. Tell me where to reopen this same task.
Compare the work with my original brief. Report each check as pass, fail, or needs review, with evidence. List missing information and actions you did not take. If the work has drifted, remove the extra scope and continue only with the approved deliverables. Do not send or publish anything.
Before stepping away: confirm cloud execution and source access. Local computer steps still need the computer available. Reopen this same task to continue. A long-running job can pause for approvals, missing inputs, or limits.
Example · BullyBearAI An overnight Codex session worked through epic #1090 and produced eight draft pull requests—proposed code changes ready for review.
The run reported 11h 25m of session elapsed time and a separate 2h 46m goal timer. It did not deploy to AWS. The Claude edition adapts the workflow; the original run used Codex.
Start in Claude Code with the project open and GitHub access available. Give it the outcome and completion checks below. Check Local or Cloud before stepping away.
In Claude Code, work through the AFK-ready features under [repository and epic]. Refresh the issues and project rules first; skip blocked or human-dependent work. Use up to two subagents on independent tasks, with separate branches. For each feature, implement it, run the relevant tests, update documentation, and open a draft pull request. Save progress after each feature. Stop after [time limit], at [usage limit], or when no ready work remains. Report any limit you cannot measure. Do not merge or deploy. Finish with PR links, test evidence, dependencies, and what needs a person.
A second chat gets a clean look at the work. Include the epic, pull requests, and the original completion checks. If deployment is part of the review, name the exact test environment.
Review BullyBearAI epic #1090 and PRs #1128, #1129, #1130, #1134, #1135, #1136, #1139, and #1140. Fetch current changes and dependencies; do not rely only on the author's summary. Check the code and rerun tests. Fix problems on the affected PR branch and push the updates. Recheck after every fix. Deploy and test sequentially only in [authorized AWS test account, region, and environment]. Stop if that target is unspecified. Record the deployed commit and results for each PR. Do not merge or touch production, broker access, or trading controls. Finish with the review order and remaining blockers.
Before you leave: check that the agent can reach the files and tools, knows when to stop, and has somewhere to save progress. Long jobs can still stop for limits, questions, or missing access.
Before you call it done
Read the document, play the game, inspect the task, or test the change. Ask what was not checked. Save the instructions that worked so your next task starts further along.