# Like My Code

> Use when a repo's code style must become written policy — harness conventions (coding.md, boundaries.md) backed by repo evidence. Triggers: "내 코드 스타일로 컨벤션 뽑아줘", "이 레포 스타일을 정책으로", "like my code", "extract conventions from this repo".

- Skill: `newkayak12/like-my-code` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx get-skill add newkayak12/like-my-code`
- Raw SKILL.md: https://api.skillmd.com/api/skills/newkayak12/like-my-code/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: MIT
- Author: newkayak12 (https://skillmd.com/u/newkayak12)
- Updated: 2026-11-14
- Page: https://skillmd.com/skills/newkayak12/like-my-code

---


## Standing Mandates

- NEVER write a rule the repo doesn't show: an unevidenced rule is the model's taste filed as the user's policy, and
  every later session obeys it. Each positive rule needs ≥2 `file:line` examples from ≥2 files and a consistency
  ratio; absence/boundary rules ("never imports", "append-only") carry a violation count or commit evidence plus one
  `file:line` example.
- A pattern below 70% consistency, or seen in fewer than 3 files, is *uncertain* — listed, not written.
- ALWAYS cite from `references/sources.md` only. No fitting source → `repo-only`. Never cite a book or item from memory.
- NEVER adopt a habit that contradicts a cited source silently — it goes under *Conflicts*; the user decides keep or drop.
- NEVER sample generated, vendored, or fixture code: its style is the generator's, not the team's. Formatter/linter
  configs in the repo outrank counted habits.
- NEVER overwrite a filled line in an existing convention file: someone wrote it deliberately. Fill `<!-- fill -->`
  placeholders and append; show the diff.
- Cite only a section whose title names the rule's topic (mapping: `references/sources.md`); a stretched fit is
  `repo-only`.
- Rules are checkable: "functions ≤ 30 lines (91% of 212)", not "functions are short".
- Every count comes from a command you ran (grep, a script), never an estimate.
- A source repo or target that can't be read is `[확인 필요: ○○]`; stop before Write.

Goal: every written rule has its count, ≥2 examples and a source or `repo-only`; everything seen but not written is
listed as uncertain or conflict; the closing tally recounts to the written files.

# Like My Code

**Not for** reviewing code against conventions (`develop:clean-code`), deciding what the architecture should be
(`develop:architecture-designer`), or installing the empty convention files (`harness:install`).

## Process

**1. Collect.** Source repo (default: current project) · optional sub-path · target dir for the policy
(default: `<current project>/.claude/conventions/`). Source ≠ target is fine — a reference project's
style can govern a new one.

**2. Sample.** Recently touched hand-written files: `git log --since=6.months --name-only`, drop
generated/vendor/test-fixture paths, keep up to ~40 files across modules. Read formatter and lint configs
(`.editorconfig`, `ktlint`, `detekt.yml`, `.eslintrc*`, `ruff.toml`, …) first — they are rules already.

**3. Observe.** Count, per dimension, what the samples do: naming, functions, errors, comments, structure,
dependencies, boundaries (what to count: `references/dimensions.md`).

**4. Cite.** For each rule ≥70%, attach the source from `references/sources.md` that states it
(`[CC ch.3]`, `[RF Long Function]`, `[CA ch.22]`, `[KT naming rules]`). A rule the source contradicts
→ *Conflicts* with both sides quoted.

**5. Show the draft.** Rules, uncertain, conflicts — the user edits before anything is written.

**6. Write.** `coding.md` gets Naming & structure / Style / Dependencies; `boundaries.md` gets its three
lists: *Do not modify without explicit user approval*, *Off-limits entirely* (generated/vendored dirs),
*Requires a dependent update when changed*. Missing file → create from the harness template headings. Report the diff.

**7. Recount.** Count the rules in each written file, those with a source, and `repo-only`; fix the tally line if the
files differ.

## Output Template

```
Source: <repo>@<sha> · N files sampled · configs: .editorconfig, detekt.yml
Target: <dir>/.claude/conventions/

coding.md
## Naming & structure
- Predicates start with is/has/can — 94% (118/125) · ex: src/a/Order.kt:42, src/b/User.kt:17 · [CC ch.2]
(then Style, Dependencies in the same shape)

boundaries.md
## Do not modify without explicit user approval
- db/migration/** — append-only (0 edits after add in 38 files) · ex: db/migration/V12__add_user.sql:1 · repo-only
(then Off-limits entirely, Requires a dependent update when changed)

Uncertain: [pattern — ratio — why not a rule]
Conflicts: [habit + ratio] vs [source says] — keep or drop?

Tally: <n> rules written (<s> sourced · <r> repo-only) · <u> uncertain · <c> conflicts awaiting your call
```

## What Claude Does / What You Do

| Claude | You |
|--------|-----|
| Samples, counts, cites only from sources.md, separates uncertain and conflicts | Names the source repo and target |
| Shows the draft before writing; fills placeholders without overwriting | Edits the draft; decides each conflict |
| Writes coding.md / boundaries.md and reports the diff | Commits the convention files |

## Related Skills

- `develop:clean-code` — reviews code against these conventions; its references explain each dimension
- `develop:clean-architecture` — the dependency rule behind the Dependencies section
- `harness:install` — scaffolds the empty convention files this skill fills
- `write:like-me` — the same idea for prose

