debriefing-code-changes— agent skill

Output Contract

Produces: a Code Change Debrief Markdown report grounded in a git diff or commit range, with: change summary, files-to-understand shortlist, per-change explanation, pattern/concept map, caveats, practical quiz, and optional durable notes.

Does NOT produce: generic CS lessons, ungrounded pattern spotting, code edits, refactors, or deploy approval.

Hands off to: the developer's learning loop — they read the debrief, answer the quiz, then decide whether to save notes, add tests, or request a separate code review/fix skill.

Process

Phase 1 — Collect the actual change set

1. Ask for or infer the target range: unstaged diff, staged diff, last commit, branch diff, or explicit commit range.
2. Prefer git facts over chat memory. Run scripts/collect_diff.py or equivalent git commands to capture: changed files, diff stats, commits, file excerpts, and touched functions/classes where available.
3. If the diff is too large, shortlist the 3–7 files with the highest learning value: core logic, new abstractions, boundary code, data model changes, tests, and security-sensitive paths.
4. If no diff or commit range exists, stop and ask for one. Do not debrief from the chat transcript alone.

Why: the lesson lives in code the developer now owns. Chat transcripts are noisy and often explain intentions that never landed in the repository.

Read references/debrief-rubric.md before writing the report.

Phase 2 — Explain intent and structure per non-trivial change

For each selected change, produce this block:

### [File or subsystem]
What changed: [plain-English summary]
Why it was shaped this way: [architecture/design rationale inferred from the code]
Key code to read: [file:function or file:line when available]
Pattern or concept: [only if visibly present in the code]
Tradeoff accepted: [what this design makes easier vs harder]
One thing to watch: [edge case, coupling risk, missing test, performance/security caveat]

Rules:

• Cite actual files/functions for every claim.
• Say "no named pattern needed" when the code is straightforward.
• Separate observed facts from plausible intent. Use "the code appears to..." when intent is inferred.
• Prefer 3–5 high-signal explanations over exhaustive file-by-file commentary.

Phase 3 — Surface caveats the developer should own

Cover the four caveat classes explicitly:

1. Edge cases — null/empty inputs, race conditions, boundary values, retries, partial failures.
2. Performance — loops over large collections, repeated I/O, N+1 queries, caching, render frequency.
3. Security/privacy — auth checks, secret handling, injection risk, unsafe deserialization, logging sensitive data.
4. Tech debt — duplicated abstractions, TODOs, unclear ownership, thin tests, temporary adapters.

If a class has no visible issue, write: No obvious issue in the inspected diff. Do not invent risk to fill the section.

Read references/concept-map.md for pattern terminology and guardrails.

Phase 4 — Build a practical quiz loop

Generate 5–8 questions that test whether the developer can reason about the actual diff:

• Explain: "Why did this change need [adapter/interface/state split]?"
• Debug variant: "What breaks if [API returns null / event arrives twice / list is empty]?"
• Tradeoff: "What did this design make easier, and what did it make harder?"
• Test placement: "Where would you add a test for [specific branch]?"
• Deployment judgment: "Which part of this diff would you be nervous to deploy, and why?"

After the user answers, grade against the diff, correct misconceptions, and ask one follow-up question on the weakest area.

Phase 5 — Save learning notes only when asked

Offer to save a concise note if the user wants the learning to compound. The note should include:

• commit/range reviewed
• concepts learned
• recurring personal blind spots
• tests or follow-up work identified
• links to files or PRs

Do not write notes, commit files, or publish anything without explicit approval.

Rules

1. Start from git diff, git show, or a commit/PR diff — never from chat transcript alone.
2. Every architecture or pattern claim needs a file/function reference.
3. Do not label code with a design pattern unless the structural evidence is present.
4. Keep the tone educational, not performative: teach the code the user shipped, not a textbook chapter.
5. If the debrief exposes a bug or risky design, recommend a separate fix/review step rather than silently editing code.
6. Include caveats across edge cases, performance, security/privacy, and tech debt; explicitly say when no obvious issue is visible.
7. Quiz questions must be practical and diff-specific.
8. Ask before saving notes or making any repository changes.

<!-- Built with Agent Engineer Master — get your own production-ready skill: www.agentengineermaster.com/skill-engineer -->