| name | enforce-coding-guideline |
| description | Walk the current branch's diff against every applicable coding guideline from documents/coding-guidelines/. Finds the rules that govern the diff (search.ts + per-rule folders), runs each rule's check.sh where one exists, checks the rest by eye, APPLIES every Strong + Medium violation without asking, verifies once, and loops internally until nothing material is left. Use after implementation feels done but before opening / re-pushing a PR. |
Enforce Coding Guideline
Mechanical rule-by-rule walk of the project's coding guidelines against the current diff. The whole point is the enumeration — you cannot decide "this rule doesn't apply" without checking the diff against it.
Not architecture — /rethink-and-simplify handles that. This skill is style + repo conventions: hardcoded colors, missing active-states, repository-pattern breaks, oversized components, nested ternaries, useEffect rules, anything captured in documents/coding-guidelines/<group>/<slug>/rule.md.
The skill loops internally. It does NOT emit a verdict line for an outer orchestrator. It returns only when the diff passes every applicable guideline.
Layout
Guidelines live one rule per folder:
documents/coding-guidelines/
search.ts # find rules by keyword
check-all.ts # run every check.sh, aggregate pass/fail
<group>/<slug>/
rule.md # the rule prose (always present)
check.sh # present ⇒ programmatic gate; absent ⇒ review by eye
Groups: frontend, backend, e2e, sharing-code.
When to invoke
- After
/rethink-and-simplify has settled the architectural shape — style fixes apply against a stable baseline that way.
- After implementation looks "merge-ready" but before the human reviewer sees it.
- After every review-iteration push during an open PR.
Procedure
Each pass is one cycle of: read diff → list applicable rules → check each one-by-one (run check.sh where it exists, eyeball the rest) → apply Strong + Medium → verify if applied → decide whether to loop.
1. Read the diff — line-by-line, with one verdict per hunk in chat
git diff origin/main..HEAD
git diff
git status --short
Note which areas the diff touches — that determines which groups apply:
packages/backend/** → backendpackages/frontend/** → frontende2e/** → e2e- shared / boundary pure TS →
sharing-code
Skimming the diff is the canonical failure mode. Even though /rethink-and-simplify ran before this, the reshape stage may have introduced new line-level bugs (accidental token swaps, off-by-ones in math that was correct pre-reshape, partial refactors). For every changed file, open it with the Read tool. For every hunk, emit ONE chat line with this structure before you touch the rule enumeration:
<file>:<line> — <one-line description> — checked: <A>/<B>/<C>
A intent: <every changed token/literal/classname/keyword listed and confirmed intentional; flag accidental token swaps like `bg-surface-1` → `bg-surface-2`>
B math: <if a constant became a function, every caller's `K - 1` / `K + 1` / `K + extras` walked — does the function's return already bake in the adjustment? off-by-ones in grid-template / nth-child / Array-length math are silent at defaults and break in non-default states>
C consistency: <if one case was "special" before and is now dynamic, EVERY consumer that referenced the old special case is updated — grep for the old name and confirm>
ANY hunk that fails A/B/C → fix it as a Strong finding before the rule walk begins. If your chat output does not contain one verdict line per hunk, you have skipped the read pass and the rule walk cannot self-terminate clean.
2. Enumerate EVERY rule — ls is non-negotiable
check-all.ts is NOT enumeration. It only runs the check.sh files — i.e.
the small subset of rules that have a programmatic gate. The MAJORITY of rules
are eye-only (no check.sh). Running check-all and stopping there means you
audited maybe 30% of the rules and skipped the rest. That is a defect.
For each touched group, you MUST ls the folder and read the FULL list of rule
folder names out loud (in chat) — every single one — before deciding which apply:
ls documents/coding-guidelines/frontend
ls documents/coding-guidelines/backend
ls documents/coding-guidelines/e2e
ls documents/coding-guidelines/sharing-code
Each folder name IS the rule's name — the slug usually tells you what the rule
is about without opening it. Walk the full list and decide, rule-by-rule,
whether it applies. State that decision in chat for every rule you mark as
applicable (slug + one-line reason). You don't have to justify rules you skip,
but you DO have to have considered each one — and the way you prove that to
yourself is by reading the full ls output before listing applicable rules.
Keyword search is a supplemental shortcut, NOT a substitute:
./documents/coding-guidelines/search.ts --group frontend --keywords "className, fetch, useEffect, as unknown"
search.ts tags each hit [check] (a check.sh exists) or [manual]. Use it to
catch extra hits AFTER you've walked the full ls, never as a replacement for
it — it only matches literal tokens and misses anything phrased differently
from the diff vocabulary.
If no groups apply, return immediately with "no applicable guidelines found."
3. Read each applicable rule.md
Use Read on the rule.md of every rule in scope. Not skim. Read the prose and the
✅/❌ examples.
Do not skip "obviously inapplicable" rules at this stage. "Obviously" is where
misses happen — the 100-line limit, the cn()-not-template-literal rule, the
no-raw-useEffect rule all read as "obvious" until they get missed.
4. Build the checklist — one item per rule folder
Use TodoWrite, one item per rule folder per touched group:
[frontend/100-line-component-limit...] — check (has check.sh)
[frontend/no-raw-useeffect] — check (manual)
[frontend/never-use-hardcoded-colors...] — check (has check.sh)
[backend/all-db-access-goes-through-a-per...] — check (manual)
...
This list IS the audit plan. You cannot skip an item because it "obviously
doesn't apply" — that's exactly the failure mode this skill exists to prevent.
5. Walk the checklist one-by-one
For each item:
- If the rule has a
check.sh, run it — it's the deterministic answer:
bash documents/coding-guidelines/<group>/<slug>/check.sh
(Or run them all at once: ./documents/coding-guidelines/check-all.ts.) A
non-zero exit prints offending file:line. Cross-reference against your diff:
a hit your change introduced is a Violation; a pre-existing hit elsewhere is
out of scope (note, don't fix).
- If there's no
check.sh (architectural / judgment rule), check by eye:
does any changed file contain the pattern the rule names? Does any changed
line introduce the rule's "❌ Wrong" example? Read the rule text, don't trust
memory.
Record per item: Violation (file:line + concrete fix, ranked Strong /
Medium / Minor) or Clean.
6. Apply Strong + Medium — without asking
Use Edit. Do not write "Want me to apply?" — the user invoked this to GET
fixes. Strong first (descending impact), then Medium. Minor: skip unless trivial.
If a fix expands scope beyond the PR's intent, note it as a follow-up and
don't apply.
7. Verify — only if fixes applied
If any Strong / Medium fixes were applied, run the project's lint + typecheck
ONCE at the end, plus re-run the relevant check.sh (or check-all.ts) to
confirm the violation is gone. If verification fails, fix forward in this pass.
If no fixes were applied, skip verify entirely.
8. Decide whether to loop
- Applied Strong / Medium fixes → loop once to confirm no new violations.
- Found nothing → stop.
Hard ceiling: 2 internal passes.
9. Final report
Groups audited: frontend, sharing-code
Rules walked: 14 (6 via check.sh, 8 by eye)
Rules with violations: 3
Pass 1 — applied:
- [frontend/never-use-hardcoded-colors...] login.tsx:42 — bg-blue-600 → bg-accent-1 (check.sh)
- [frontend/interactive-elements-must-have-four...] Button.tsx:38 — added active:scale-95 (eye)
- [frontend/100-line-component-limit...] ChatPanel.tsx — split into ChatPanel + MessageList (check.sh)
Pass 2 — clean.
Verify: lint clean, types clean, check-all clean. (or "skipped — no fixes applied")
No VERDICT line. The skill has already self-terminated.
Hard rules
ls every applicable group, read the FULL list, then decide. Skipping the ls and just running check-all.ts is the #1 failure mode of this skill. check-all covers only the check.sh-gated minority; the eye-only majority gets silently skipped. If your audit report doesn't begin by quoting the ls output (or naming every rule folder in every touched group), the audit didn't happen.- Run the check.sh, don't eyeball a checkable rule. If a rule has a check.sh, its verdict is the script's exit code — running it is non-negotiable.
- No diplomacy. "This violates the hardcoded-color rule." Soft critiques are invisible critiques.
- No vague findings. Every finding cites file:line + the rule slug + the concrete fix.
- Don't expand the PR. Out-of-scope / pre-existing findings → mark as follow-up, don't apply.
- Don't second-guess the rule. The rule is the rule.
- Stop in finite time. Cap is 2 passes.
Anti-patterns
- Don't run
check-all and call it enumeration. check-all runs only the programmatic gates — a minority slice. Eye-only rules (majority of the catalog) get silently skipped. ls the group folders, read the full list, decide rule-by-rule.
- Don't skim and pattern-match. "I looked, nothing jumped out" is the failure mode this skill exists to prevent. The TodoWrite enumeration is non-negotiable.
- Don't ask "should I apply?" The user invoked the skill to get fixes.
- Don't verify per-fix. One verify at the end of a pass.
- Don't re-find what you already fixed. Each pass is on the current state.
- Don't audit groups that don't apply. A backend-only diff doesn't need a frontend walk. The scope filter in step 1 is real.
Worked example
Pass 1 on a branch touching packages/frontend/src/components/login.tsx:
Groups in scope: frontend only.
Checklist (from ls documents/coding-guidelines/frontend + search.ts):
[frontend/100-line-component-limit...] — check.sh → login.tsx 87 lines → clean
[frontend/no-raw-useeffect] — eye → no useEffect in diff → clean
[frontend/never-use-hardcoded-colors...] — check.sh → bg-blue-600 line 42 → VIOLATION (Strong)
[frontend/interactive-elements-must-have...] — eye → new ghost <Button> line 38, no active: → VIOLATION (Medium)
[frontend/use-cn-never-template-literals...] — check.sh → clean
... (rest of frontend rules, all clean)
Applied 2 fixes. Verify clean. → Loop once. Pass 2: all clean. Stop. Done.
