Debrief

Overview

A bottom-up friction-capture skill. At the end of a Claude Code session, scan the conversation already in context, identify project-specific friction (project rules and project knowledge), and propose CLAUDE.md edits the engineer can approve. Bias toward zero proposals — returning "session looks clean" is a first-class outcome, not a fallback.

This skill is part of the future Orbitant orbitant-engineering plugin's bottom-up half. It is decoupled from /ground-control (the top-down half, not yet built).

When to Use

• The engineer types /debrief (no args) at session end.
• The engineer types /debrief <hint> to focus on a specific area (e.g. /debrief the API naming convention thing).
• Skip with the message "No active session content to debrief — run me at the end of a session that had real engineering work." if the session has no engineering work in it (e.g. a brand-new chat with only this /debrief command).

Prerequisites

• The skill assumes Claude has the full Claude Code session already in context. It does not re-read the session from disk.
• The project may or may not have a CLAUDE.md. The skill handles both cases (see the application step).

Process

1. Read the existing CLAUDE.md

Before proposing anything, read the project's CLAUDE.md if it exists. Hold its contents in mind so you can dedupe — never propose something already covered.

If the project has subdirectory CLAUDE.md files (e.g. api/CLAUDE.md, mobile/CLAUDE.md), read those too. They count for dedupe.

If no CLAUDE.md exists at all, that's fine — proceed. The application step (later) will offer to create one.

2. Scan the session for friction

Look for two narrow categories. Nothing else qualifies.

Category A — Project rules. Explicit corrections that signal a project-wide convention.

Examples:

• "we don't use moment.js"
• "snake_case for API fields, not camelCase"
• "don't add try/catch around X"
• "we always write tests in Vitest, not Jest"

Category B — Project knowledge. Facts about the codebase Claude lacked, that the engineer had to volunteer.

Examples:

• "the API is at /api/v2/"
• "auth tokens live in the tokens table"
• "run npm run gen after editing schemas"
• "the dev server requires DATABASE_URL in .env.local"

3. Apply the bar

For each candidate, ask: Would adding this to CLAUDE.md have changed Claude's behavior in this session? If no, drop it. The bar is single, mechanical, evidence-driven. There is no "must see N times" threshold — the principle handles it.

4. Apply the exclusions

Drop any candidate that fits any of these:

• Engineer changed their mind mid-session. ("Actually, let's not do X." — that's a reversal, not a rule.)
• One-off bugs Claude fixed after one prompt. (Wouldn't have been prevented by CLAUDE.md.)
• Style nudges Claude got right after one correction. (Already responsive — no future risk.)
• General coding wisdom. ("Don't use var" is true everywhere; CLAUDE.md is for project-specifics.)
• Anything already covered in existing CLAUDE.md (deduped in Step 1).

5. Skill-execution friction counts the same

If a correction surfaced during a skill invocation (e.g., the engineer corrected you while running /twd or another skill), it counts the same as any other correction. Do not discount it as "just clarifying the skill's instructions". The bar is still: would it have changed your behavior if it had been in CLAUDE.md? If yes, it qualifies.

6. Require evidence per surviving candidate

Every surviving candidate must have at least one quoted user message or a concrete missing-knowledge moment from the session. No evidence → no proposal. This makes the empty-result outcome mechanical, not a judgment call.

7. Bias toward zero

If you find nothing that meets the bar, that is the correct outcome. Do not invent proposals to feel productive. A clean session is a successful one. Returning zero proposals is celebrated, not apologized for.

8. Output the result

When proposals exist

Return ONE message in this exact shape — and nothing else:

Found N moments worth capturing.

## 1. <short title>
**Type:** project rule | project knowledge
**Evidence:** "<quoted user message or specific session moment>"
**Proposed CLAUDE.md addition:**

> <exact markdown that would land in CLAUDE.md>

## 2. ...

---
Tell me which to apply: "all", "1 and 3", "1 with edit: <change>", or "none".

Each proposal must include the four labelled fields:

• **Type:** — exactly project rule or project knowledge.
• **Evidence:** — at least one quoted user message OR a concrete description of the missing-knowledge moment. The engineer must be able to recognize the moment without scrolling.
• **Proposed CLAUDE.md addition:** followed by a blockquoted markdown block — the literal text that would land in CLAUDE.md, not a description of it. The engineer reads this verbatim, so it must be ready-to-paste.
• The --- separator and the Tell me which to apply: line at the very end.

When nothing surfaces (the celebrated path)

Return this message — exactly this, no embellishment, no additional commentary:

Session looks clean. No project rules or knowledge gaps surfaced — your CLAUDE.md held up.

If something nagged you that I missed, run `/debrief <hint>` (e.g. `/debrief the import ordering thing`) and I'll focus there.

This is the default outcome the skill is biased toward. Do not append "but here's something just in case…" or hedge — a clean session is the good outcome.

9. Apply selected proposals

Wait for the engineer's reply. Once they answer:

"none" or no selection

Reply: "OK, nothing applied. CLAUDE.md unchanged." and stop.

"all", "1 and 3", or specific indices

For each selected proposal, write it to the project root CLAUDE.md:

1. If CLAUDE.md does not exist: Create it with a minimal scaffold:

   # <repo-name>
   
   ## <Section name from the proposal type>
   
   - <proposal markdown>
   

   Use the repo directory name as the H1. Pick the section heading based on the proposal type — e.g. ## Conventions for project rules, ## Project knowledge for project knowledge.

2. If CLAUDE.md exists:

   • If the proposal clearly fits an existing section heading (e.g., a new convention rule fits an existing ## Conventions section), append within that section.
   • Otherwise, append a new section at the bottom of the file.
   • Never modify or delete existing content. Append-only.

3. Do not target subdirectory CLAUDE.md files (api/CLAUDE.md, etc.) — always edit project root in v1. The engineer can move content manually if needed.

4. Do not stage and do not commit. After writing, report:

   Applied [list of proposal titles] to CLAUDE.md.
   Review the diff and commit when ready.
   

"1 with edit: "

Apply the engineer's edit verbatim. Do not second-guess. If the requested change is genuinely unclear, ask one clarifying question, then apply.

Engineer pushes back ("that's not actually a rule, I just changed my mind")

Drop or revise the proposal based on the engineer's reasoning. Do not defend the proposal. The engineer is the source of truth.

Edge cases

• Hint matches nothing. If /debrief <hint> was used but no friction matches the hint, reply:

  I scanned for "<hint>" but couldn't find friction matching it. Either the topic didn't surface in this session, or your existing CLAUDE.md already covers it.

  Do not fabricate a match.

• Engineer rejects all proposals. Reply "OK, nothing applied. CLAUDE.md unchanged." Do not re-pitch.

• Multiple /debrief calls in one session. This skill is stateless. Re-running scans the session afresh. Already-applied items will not re-surface (caught by the dedupe pass against CLAUDE.md). Previously-rejected items may re-surface — that's correct: the engineer can change their mind.

• Brand-new fresh repo with no CLAUDE.md. The application step handles this (creates the file with a minimal scaffold). No special detection logic needed.

• Very short session. Run normally. If the bar isn't met, return the "session looks clean" message.

• Session crossed many topics. Trust the bar to filter. The per-proposal evidence requirement keeps quality up regardless of session breadth. Do not summarize every topic — surface only what meets the bar.

Important

• Bias toward zero. Your default outcome is "session looks clean". Finding something is the exception that needs justification, not the rule.
• Append-only. Never modify or delete existing CLAUDE.md content.
• Engineer is source of truth. If the engineer pushes back on a proposal, drop or revise — never defend.
• No staging, no commits. The engineer owns the git flow.
• No proposals beyond CLAUDE.md edits. This skill does not propose new skills, subagents, hooks, slash commands, or settings. If the engineer asks, point out that those are deliberate-design artifacts and out of scope for /debrief.