Rules Curator

Rules Curator

Overview

Rules Curator turns conversation-level intent into durable repository guidance. Its job is curation: classify first, write only when the candidate belongs in a high-authority rule file.

Formerly rules-keeper.

Workflow

1. Extract the candidate rule.

• Capture the smallest actionable sentence.
• Remove one-off task context, implementation narration, and temporary preference wording.
• Completion criterion: the candidate can stand alone as future-agent guidance.

2. Classify the destination.

Completion criterion: state the chosen destination and why it fits before writing.

3. Apply the root-rule gate.

Write to root instruction files only when all are true:

• The rule is expected to remain valid across future sessions.
• It applies across tasks, modules, pages, or agents.
• It tells future agents what to do or avoid.
• The user explicitly wants durable behavior, or confirms it after review.
• It is concise enough for a root instruction file, or has a concise root summary with a documentation link.
• It does not duplicate or conflict with an existing rule.

If any item fails, route to the better destination instead of forcing a root rule.

4. Ask before high-authority writes when confirmation is missing.

• If request_user_input is available, ask one question with 2-3 choices.
• Put the recommended option first and suffix its label with (Recommended).
• Stop until the user answers.
• If that tool is unavailable, ask in chat with a numbered list and wait for the reply.

Recommended choices:

1. Standard doc plus root summary, for longer rules.

2. Root rule, for short mandatory guidance.

3. Learning or decision record, for non-mandatory knowledge.

4. Write the rule.

• Prefer scripts/update_dynamic_rules.py.
• Root file is AGENTS.md.
• Mirror to existing claude.md / CLAUDE.md / gemini.md / GEMINI.md by default.
• Create missing mirror files only when the repository or user explicitly wants mirrored agent files.

6. Report the result.

• Destination chosen.
• Exact rule text written or reason no write happened.
• Files changed.
• Duplicate or conflict notes.

Script Usage

Append a root rule:

python "C:\path\to\skills\rules-curator\scripts\update_dynamic_rules.py" --repo-root . --rule "All runtime logs must be English ASCII" --scope "全仓库"

Preview without writing:

python "C:\path\to\skills\rules-curator\scripts\update_dynamic_rules.py" --repo-root . --rule "All runtime logs must be English ASCII" --dry-run

Sync existing mirror files:

python "C:\path\to\skills\rules-curator\scripts\sync_agents.py" --repo-root .

Create missing mirror files only when explicitly requested:

python "C:\path\to\skills\rules-curator\scripts\sync_agents.py" --repo-root . --create-missing

Rule Table

Root rules are appended under:

## 动态更新规则
| 日期 | 规则 | 适用范围 | 来源 | 备注 |
|---|---|---|---|---|

Rows use:

| 日期 | 规则 | 适用范围 | 来源 | 备注 |

Use the note field for replacements, for example supersedes: 2026-07-01 old summary.

Common Mistakes

• Do not write a clear sentence just because it is clear. Clear one-off instructions are still one-off.
• Do not append near-duplicates. Merge, replace, or mark supersession.
• Do not put long rationale in root files. Link to a standard, ADR, or solution doc.
• Do not silently create every possible agent file in a repository that only uses one.
• Do not treat a learning as a rule. Rules are mandatory; learnings are guidance.