Rule writing

Skill BCGen/skills/skills/rule-writing

Teach your coding agent — Claude Code, Cursor, or any AGENTS.md agent — your project's conventions, and turn every correction into a team-shared, git-tracked lesson.

Install

npx -y skills add BCGen/skills --skill rule-writing

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Creates, edits, or removes a single project rule in the agent's native rule format, enforcing a non-discoverability admission filter and a hard rule budget with provenance stamping. Use when the user asks to add, change, or delete a project rule, or when another skill hands over a drafted rule.

SKILL.md

3.0 KB, as published. Nobody here has run it

Rule Writing

The single write path for rule files. One invocation = one rule added, edited, or removed. Full format details, budgets, and dialect mapping live in references/rule-format-spec.md — read it before writing.

Step 1 — Understand the request

Collect: the rule text (draft from the user or a calling skill), its scope (project-wide or specific paths), and its evidence. If the intent is unclear, ask one question with a recommended answer.

Step 2 — Admission filter

Reject the draft, stating which test failed, unless it is at least one of:

Non-discoverable from the repo (landmine, tribal convention), or
Mistake-correcting with observed evidence.

Hard rejections: restates linter/formatter/typechecker config; restates something readable from the code; vague virtue ("write clean code"). Offer to rewrite a rejected draft into a passing, verifiable form when possible.

Step 3 — Detect the target

Look for .claude/ (Claude Code), .cursor/ (Cursor), AGENTS.md (fallback). Zero or multiple candidates → ask one question, recommending AGENTS.md as the portable default. Emit the dialect per the spec.

Step 4 — Enforce the budget

Compute resident usage per the spec (entry file + unscoped rules; content lines only). Then:

< 80% → proceed.
≥ 80% → warn with the numbers; prefer path-scoping this rule if its subject maps to specific files.
= 100% or would exceed → refuse, list removal/merge candidates (stale provenance, overlapping subjects), and only write if the user removes or merges one in the same operation.

Step 5 — Write (itemized only)

Add: create the rule file (or one ### entry inside the AGENTS.md managed block). Never touch neighboring content.
Edit / remove: modify or delete exactly the named rule.
Stamp the provenance comment (date · task · evidence · via).
Rules ≤ 20 content lines; split or tighten anything longer.
Never rewrite a whole file or block wholesale.

Step 6 — Report

One short confirmation: rule path, scope (resident or path-scoped), and budget usage after the write (e.g. "resident 92/150 lines"). If invoked by another skill (e.g. retro), also return the rule path so the caller can update its records.

Mistakes to refuse

Request	Response
"Add these 5 rules at once"	One rule per invocation — propose an order
"Just append it, skip the filter"	Filter is non-negotiable; explain why
"Rewrite the rules file to be cleaner"	Itemized edits only; offer per-rule merges
A rule duplicating an existing one	Point to the existing rule; offer an edit instead