Mimic

Mimic — Coding-style extraction & reproduction

What this produces

A style-profile.md: a set of decisions (not prose descriptions) an author makes, each backed by real examples mined from the repo, with do/avoid pairs and a confidence score. The profile is written to be consumed as a skill or CLAUDE.md fragment so future code generation reproduces the style. The skill ends with a hold-out validation loop that measures reproduction fidelity — do not skip it.

Core principle

Style worth extracting = the residual after auto-formatting. Strip everything Prettier/ESLint/EditorConfig already enforce (whitespace, quotes, semicolons, import order, naming case). That is reproducible by tooling and is pure noise. Target only the decisions no linter can impose: decomposition, type modeling, error handling, async patterns, test philosophy, abstraction level, and where business logic lives.

A second principle: a rule with no concrete example is worthless. Abstract rules ("prefer immutability") do not reproduce anything. A before/after pulled from the actual repo does. Every entry in the profile must cite path:line.

Inputs — collect before starting

Confirm these with the user (don't guess):

1. Author — git author name and/or email pattern (substring match, e.g. Stephane or stephane@).
2. Scope — path glob to constrain analysis (e.g. packages/feature-views/). Defaults to repo root.
3. Window — how far back (default 6 months; style drifts, recency matters).
4. Target — where to write the profile and in what form (standalone skill vs CLAUDE.md section).

Stage 1 — Select the corpus (deterministic, scripted)

Run the bundled script. It does NOT dump everything — it computes per-file line ownership via git blame (after a cheap author pre-filter to stay tractable on large repos) and keeps only files where the author dominates. This is what keeps Stage 3 inside the context budget.

chmod +x scripts/select-corpus.sh
scripts/select-corpus.sh -a "<author-pattern>" -s "<path-scope>" -d "6 months ago" -t 50 -n 40

Outputs to .mimic-corpus/:

• owned-files.txt — files the author owns ≥ threshold %, ranked. Highest-signal: read these in full.
• ownership.tsv — full ranking (pct, lines, total, path) for inspection.
• recent.diff — their diffs over the window, lockfiles/dist/snapshots excluded, capped.
• commits.txt — sha|date|subject, reveals commit granularity & discipline.

Stage 2 — Signal hierarchy

Read sources in this priority. Spend the context budget top-down.

Optional but high-value — pull the author's review comments (preferences they impose on others):

gh pr list --reviewer "@<login>" --state merged --limit 50 --json url -q '.[].url' \
  | while read url; do gh pr view "$url" --comments --json comments \
      -q '.comments[] | select(.author.login=="<login>") | .body'; done

Stage 3 — Extract the profile

This is judgment work. Do it yourself, or spawn a subagent with a clean context window if the corpus is large. For each dimension below, emit: a rule, 2 evidence citations (path:line

• one-line note), 1 counter-example (what the author demonstrably avoids), and a confidence (high/med/low, with n = number of supporting instances). Drop any dimension with < 2 instances — say "insufficient signal" rather than inventing a rule.

Dimensions:

1. Module & file decomposition — one responsibility per file? co-location of types/tests? barrel files?
2. Type modeling — discriminated unions vs classes, branded/nominal types, richness vs pragmatic any/unknown, inference vs explicit annotations.
3. Error handling & boundaries — throw vs Result/Either, where errors surface, custom error types, exhaustiveness.
4. Async & concurrency — async/await vs promise chains, parallelism patterns, cancellation, retry/backoff.
5. State & data flow — where business logic lives (component / hook / service / domain), mutation posture, dependency injection style.
6. Function shape & abstraction — arg count, options-object threshold, early-return vs nesting, when they extract vs inline, tolerance for abstraction.
7. Testing philosophy — test structure (AAA, nesting), mock posture (heavy mocks vs real deps), what gets tested vs skipped, fixture style.
8. Comments & docs — JSDoc presence, comment density, what they comment (why vs what), TODO conventions.
9. Dependency posture — build vs buy, recurring library choices, reluctance/eagerness to add deps.
10. Domain naming — semantic naming beyond lint rules (domain vocabulary, abbreviation tolerance, boolean/handler naming patterns).

Output schema per dimension:

### <dimension>
- **Rule:** <decision in one sentence>
- **Evidence:** `path/to/file.ts:42` — <why this shows it>; `other.ts:88` — <…>
- **Avoids:** <the alternative they demonstrably reject>
- **Confidence:** med (n=4)

Stage 4 — Validate (hold-out, mandatory)

A profile that isn't validated is decoration. Run an eval-style loop:

1. Pick a recent commit/PR by the author NOT used in extraction (check it's not among the cited files).
2. Reconstruct the pre-state: git checkout <sha>~1 (or a worktree).
3. Hand Claude the task + the style-profile.md, have it implement.
4. git diff the generated code against the author's real <sha> implementation.
5. Each meaningful divergence is a verdict: a missing rule (add it) or an over-rigid rule (loosen it).
6. Iterate until divergences are stylistically negligible. Report a fidelity summary (what matches, what doesn't, residual gaps).

Restore state afterward: git checkout - / remove the worktree.

Output artifact

Write style-profile.md with: a header (author, scope, window, corpus size — n files / n commits), the 10 dimensions, and a closing validation report. If the target is a reusable skill, wrap it with frontmatter so it loads contextually; if CLAUDE.md, prepend a one-line "When writing in <scope>, follow these conventions:" anchor.

Guardrails

• Formatting noise — if any rule could be enforced by Prettier/ESLint, delete it; the extraction missed the point.
• Squash merges destroy blame granularity. If the repo squashes, derive incremental style from PR diffs via gh, not git blame.
• Attribution pollution — vendored, generated, or copy-pasted code attributed to the author skews the profile. The script excludes *.lock/dist/*.snap; flag anything else that looks non-authored.
• Team vs personal — what the author does under repo constraint ≠ their preference. Review comments disambiguate; weight them.
• Sample size — below ~5 owned files or ~15 commits, label the whole profile low-confidence and say so explicitly.
• No fabrication — a dimension without ≥ 2 real examples is reported as "insufficient signal," never as an invented rule.