This repo scaffolds a Claude Code–based implementation system, adapted from best practices in the Nov 7 Office Hours.
Use this template to kick off any AI-assisted coding project using Claude Code + GitHub + Linear + Markdown.
This document outlines a repeatable Claude Code-based implementation workflow, informed by Helen's system and adapted for build-in-public or solo dev environments. It assumes use of:
- Claude Code (web and/or desktop)
- GitHub repo for all code + documentation
- Linear for milestone/sprint tracking
- Optional: Airtable for structured data tracking (vs Notion MCPs)
- Codify everything: Claude has no memory. Document all decisions and context in markdown files.
- Chunk your work: Smaller tasks = better Claude performance. Don’t ask for full apps in one go.
- Ground your agents: Keep your PRD/vision doc in context at all times.
/.github/ # Optional GitHub Actions
/claude/ # Implementation plan + learnings from Claude
implementation-plan.md
learnings/
/commands/ # Reusable Claude commands like /implement
/docs/ # Vision, setup instructions, etc.
product-vision.md
setup-instructions.md
/skills/ # Reusable Claude skills (e.g., refactor, frontend)
/scripts/ # Deployment, migrations, etc.
/src/ # App logic or script code
.gitignore # Ignore system & build files
README.md # Project overview and folder guide
package.json # (Empty or default for now)
- Use Claude Code to generate
/docs/product-vision.md– this is your PRD. - Use Claude Code to generate
/claude/implementation-plan.md. - Track any Claude insights or decisions in
/claude/learnings/. - Store custom commands in
/commands/, skills in/skills/. - Use Linear for polish tasks if needed.
🧠 Pro Tip: Claude has no memory — keep everything important in Markdown files in this repo.
-
Write a clear product doc.
-
Save to
/docs/product-vision.md -
Claude prompt:
Anytime you are lost, refer to the product vision doc.
- Ask Claude to break down the project into logical stages.
- Use voice like: “Imagine a junior dev will follow this. Be overly explicit.”
- Save plan in
/claude/implementation-plan.md - Review and revise manually.
-
Anytime Claude gives an interesting prototype or insight:
- Save the transcript snippet in
/learnings/ - Include your interpretation/decision.
- Save the transcript snippet in
- Claude reads
implementation-plan.md - Identifies next incomplete task
- Begins implementation
- Updates status (in-progress ✅ / complete ✅)
- Commits code when complete
/implement→ pick up where you left off/context→ check how much memory is used/commit→ stage, format, lint, build, commit/clear→ reset context after big commits
- Desktop: Better for complex builds, local file access
- Web: Better for PRs, fast bugfixes, repo-linked sessions
- Use
implementation-plan.mdfor big blocks of features - Use Linear for fine polish (e.g., thumbnails, CSS bugs)
- Don’t ask Claude to change actual DBs.
- Instead, create SQL migrations (Drizzle ORM, Prisma)
- Use mock data for 80% of build; only formalize DB schema when confident
- Commit frequently to avoid Claude’s context compacting
- Compacting = very slow, loses flow/state
- Use
/commitand write atomic commits