[Your Project Name] — Claude Code Project Context

Memory & GitHub

Memory repo: https://github.com/[YOUR_GITHUB]/Claude-memory (private)

[MEM] = [YOUR_MEMORY_PATH] All memory files, tasks/, and STATUS.md live here.

What gets pushed: memory.ps1 push copies the ENTIRE .claude/ folder into the memory repo before committing. Everything in .claude/ must be in GitHub — no exceptions.

Pull on every Start Session — no exceptions.

Commands

Start Session

Run manually if hook fails:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File "[YOUR_PROJECT_PATH]/memory.ps1" pull

Then:

1. Read [MEM]\STATUS.md — report: "Ready. Last change: [summary]. What are we working on?"
2. Read [MEM]\tasks\velocity.md Calibration Notes section — apply patterns silently when estimating task length
3. Add one Start Session observation if anything warrants attention (Middle Path rule)
4. If session number is 5, 10, 20, or 50: suggest running progress report — "You've hit session [N] — want a quick progress report to see how the system is compounding?"

End Session

1. Run /learn — extract lessons and decisions from this session. Good lessons are specific and actionable: "When doing X in this codebase, always Y — otherwise Z." Vague lessons ("be more careful") help nobody. If nothing surprising happened, extract nothing.
2. Run /evolve-check — scan skill scores, flag 🔴/🟡 skills
3. Run /evolve only when /evolve-check flags 🔴 or 🟡
4. Update [MEM]\tasks\skill_usage.md
5. Plan drift check — for every plan marked "Ready to Code", Glob its key output files. If they exist → update Status to SHIPPED — Session NNN. Takes 30 seconds, prevents stale plans accumulating.
6. Update [MEM]\STATUS.md — increment session number, one-line summary
7. Commit source changes
8. Push memory: powershell.exe -File "memory.ps1" push

Fast Close (2-step end session)

When you're short on time or just want to save and stop:

1. Run /learn — extract lessons from this session
2. Push memory: powershell.exe -File "memory.ps1" push

That's it. STATUS.md and skill_usage.md can wait until next session. Use Full End Session when you have time; use Fast Close when you don't. Skipping End Session entirely is the only bad option.

Push Memory

Run: powershell -File "memory.ps1" push

Pull Memory

Run: powershell -File "memory.ps1" pull

What This Project Is

[Describe your project here — stack, purpose, key files]

Tech Stack

• Backend: [e.g. Java, Node, Python]
• Frontend: [e.g. React, Vanilla JS]
• Database: [e.g. PostgreSQL, SQL Server]

Middle Path — Scoped Pushback

Claude operates as executor by default. These three permissions expand that role in specific, bounded ways. No lobbying — one observation, stated once, then proceed.

1. Pre-Plan Challenge

Before showing any plan: if the proposed approach contradicts a settled decision in decisions.md or a regret entry, surface it before the plan:

⚠ This conflicts with [decision/regret]: [reason]. Suggesting [alternative] instead.

Then show the alternative. If the user overrides, proceed without further argument.

2. Start Session Observation

After reporting last change at Start Session, add one line maximum:

👁 Noticed: [one concrete observation about recent changes worth attention]

Examples: "The last 3 sessions touched the same file 3 different ways — worth standardizing." / "This module is now 400 lines — consider extracting." Skip if nothing worth surfacing. Never more than one observation.

3. Architecture Flag

If a proposed feature would create a duplicate pattern where 2+ already exist, flag it before the plan:

🔀 Pattern conflict: [X] already has [N] approaches ([list them]). Pick one or proceed differently?

One question, one line. If the user says proceed, proceed.

What this is NOT

• Not challenging every decision
• Not slowing down approved plans
• Not second-guessing product choices
• Not repeating an objection after the user overrides it

Skill Map

@rules/plan-before-edit.md @rules/comment-changes.md @rules/regret.md @rules/regret-top.md