Claude md capture

CLAUDE.md Capture

Current Targets

echo "--- Writable CLAUDE.md files ---"
{ find . -maxdepth 6 -name CLAUDE.md -not -path "./.git/*" -not -path "*/node_modules/*" 2>/dev/null; } | sort -u
echo "--- Personal (gitignored) ---"
test -f .claude.local.md && echo ".claude.local.md (present)" || echo ".claude.local.md (absent — create for personal-only notes)"
echo "--- Line budget of root CLAUDE.md ---"
test -f CLAUDE.md && echo "CLAUDE.md: $(wc -l < CLAUDE.md) lines (root cap 120)" || echo "(no root CLAUDE.md — run /mtk-setup first)"

Overview

A session often surfaces context that CLAUDE.md was missing — a build command you had to discover, a gotcha you hit, an env quirk, a pattern the codebase follows. That knowledge should compound into project memory so the next session starts where this one ended. This skill reflects on the finished session, drafts minimal additions, shows them as diffs, and applies only what the engineer approves.

It is distinct from its two siblings:

• claude-md-audit re-grades what is already in CLAUDE.md against a rubric and fixes rot. Capture adds new knowledge this session produced.
• correction-capture records engineer corrections as reusable lessons in lessons.md. Capture records project facts (commands, gotchas, env) into CLAUDE.md — the prompt every session loads.

It honors S1.5 (CLAUDE.md is protected): append-only by default, Edit never Write, no additions applied before approval.

When To Use

• At the end of a session that revealed context CLAUDE.md was missing
• When the engineer says "save what we learned", "update CLAUDE.md", "remember this for next time", or invokes the equivalent end-of-session capture
• After discovering a non-obvious command, env requirement, or codebase pattern that future sessions will need

When NOT To Use

• First-time repo setup — use /mtk-setup (runs setup-bootstrap)
• Mid-task — capture at a natural stopping point, not while work is in flight
• Re-grading or fixing stale content — use claude-md-audit
• Recording an engineer correction — use correction-capture (writes a lesson, not a CLAUDE.md fact)
• The session surfaced nothing durable — say so and stop. Do not manufacture additions to look useful.

Workflow

Phase 1 — Reflect

Review the session. What context, had it been in CLAUDE.md at the start, would have saved time? Look specifically for:

• Commands discovered or corrected (build, test, dev, lint, deploy, migrate)
• Gotchas — ordering dependencies, prerequisites, "delete X if Y" recovery steps
• Environment / config quirks — required vars, flags, version pins
• Codebase patterns that were not obvious from a first read

If none of these came up, stop here and tell the engineer there is nothing worth capturing.

Phase 2 — Decide destination (team vs personal)

Destination	When
CLAUDE.md (committed, team-shared)	Project facts every contributor and agent benefits from
.claude.local.md (gitignored, personal)	Your own preferences, local-only paths, machine-specific setup

Heuristic: first-person preference ("I like to run…") → personal. Project fact ("tests must run with --runInBand due to shared DB state") → team. When in doubt, propose it as personal.

For monorepos, prefer the nearest package CLAUDE.md when the fact is local to one package; only put cross-cutting facts in the root file.

Phase 3 — Check for duplicates

Before drafting, grep the target file (and tasks/lessons.md) for the keyword. If the fact is already documented, do not re-add it — update the existing line only if it is now wrong.

grep -in "<keyword>" CLAUDE.md .claude.local.md tasks/lessons.md 2>/dev/null

Phase 4 — Draft additions (concise)

One line per concept. CLAUDE.md is part of the prompt — brevity is the whole point. Format: `<command or pattern>` — <brief description> and match the existing file's style (table vs bullet list — do not impose a new structure).

Avoid: verbose explanations, restating what the code already says, generic best practices, and one-off fixes unlikely to recur.

Mind the budget: the root CLAUDE.md cap is 120 lines. If an addition would push it over, propose moving detail to the relevant .claude/rules/ file instead, or drop the lowest-value candidate. The proposal must also pass the relative gate — bash scripts/growth-gate.sh CLAUDE.md <proposed> — because a capture pass that grows the file a few percent every session ratchets the always-loaded context until the absolute cap finally trips.

Phase 5 — Show proposed changes (no edits yet)

Stop before any edit. For each candidate, show where it goes, the diff, and one line of why it helps future sessions.

### Addition 1 → ./CLAUDE.md (Gotchas section)

**Why:** Tests silently corrupt each other without --runInBand; future sessions
would re-discover this the hard way.

\`\`\`diff
+ Tests must run sequentially (`pytest --runInBand`) — shared DB state.
\`\`\`

Then ask: "Apply these? (yes / partial / no)"

Phase 6 — Apply with approval

After explicit approval, apply with Edit (never Write). Append into the relevant section; never restructure. If the engineer approved "partial", apply only the named subset. If a personal item was chosen and .claude.local.md does not exist, create it (it is gitignored by setup-bootstrap — verify it is listed in .gitignore before creating; add it if missing).

Rules

• Never apply before approval. Phase 5 output is a proposal, not an edit.
• Append over rewrite. New facts go at the end of the relevant section.
• Use Edit, never Write on CLAUDE.md (S1.5 protected-file rule).
• One line per concept. If you need a paragraph, it belongs in .claude/rules/.
• Match the file's existing style. Do not impose a template.
• No manufactured additions. "Nothing to capture" is a valid, common outcome.

Common Rationalizations

See .claude/skills/context-engineering/SKILL.md for the shared MTK table. Capture-specific traps:

• "I should add a few things so the session feels productive." No. Capture only what was genuinely surfaced. Empty output beats noise in the prompt.
• "This gotcha is obvious." If you had to discover it this session, it was not obvious to a cold-start agent — capture it.
• "I'll rewrite the Commands section to be cleaner while I'm here." No. That is claude-md-audit's job, gated on its own rubric. Capture appends; it does not refactor.
• "This is a personal preference, but the team file is easier." Default personal. Promotion to the team file is the engineer's explicit call.

Red Flags

• Additions applied before the engineer approved them
• A whole-section rewrite proposed instead of an append
• Generic best practices ("write tests", "use good names") added as if project-specific
• Root CLAUDE.md pushed over 120 lines by the additions
• .claude.local.md content committed (it must stay gitignored)

Verification

• Phase 1 reflection named concrete session learnings (or "nothing to capture")
• Each addition is project-specific and one line per concept
• Duplicates were grep-checked before drafting
• Proposal shown as diffs with a one-line "why"; no edits before approval
• Edits used Edit with explicit old/new strings (never Write)
• Root CLAUDE.md still within its 120-line budget after additions
• Personal items went to .claude.local.md, not the committed file