Response style sync

Response Style Sync

Overview

Use this skill to manage project rules as reusable JSON profiles and sync them into AGENT.md / CLAUDE.md. The skill supports three separate policy layers:

• response style
• agent execution policy
• LLM / proxy usage policy

Quick workflow

1. Read any existing project rule sources.

• Check user instructions from the conversation.
• Read existing AGENT.md, CLAUDE.md, or other rule files if they exist.
• Extract stable defaults, not one-off task requests.

2. Create or update the three profiles.

• response-style.profile.json
• agent-policy.profile.json
• llm-policy.profile.json
• Recommended path: <project>/.codex/rules/

3. Sync the project files.

• Run scripts/sync_project_rules.py --project <path> --style-profile ... --agent-policy ... --llm-policy ...
• By default this updates both AGENT.md and CLAUDE.md.
• The script preserves content outside the managed block when files already exist.

4. Keep rules operational.

• Write short rules another agent can follow.
• Prefer explicit behavior over slogans.
• Keep style, workflow, and LLM-cost policy separate.

Key naming convention

Use the new shorter keys by default:

• profile_id
• language_mode
• detail_level
• planning_rules
• cloud_api_rules

The script still accepts older aliases such as profile_name, language_policy, or planning_policy for backward compatibility.

Profile split

response-style profile

Use for:

• Korean-first or match-user language behavior
• answer structure
• detail level and tone
• verification reporting style
• formatting preferences

agent-policy profile

Use for:

• ask-before-proceed rules
• dev-plan-first workflow
• parallel-agent preference
• scope expansion restrictions
• persona and review stance

llm-policy profile

Use for:

• cloud API restriction policy
• proxy-first rules
• preferred repos such as multi_model_tui
• latest OSS proxy research checklist
• cost / compatibility / compliance notes

Managed-file policy

The sync script uses a managed block:

• <!-- BEGIN RESPONSE-STYLE-SYNC -->
• <!-- END RESPONSE-STYLE-SYNC -->

Keep generated rules inside this block. Keep project-specific exceptions outside it.

Recommended extras to propose

If the user asks for a project rule system, also suggest these:

1. One profile bundle per repo

• Keep the three JSON files under .codex/rules/.

2. Presets

• Add concise, standard, detailed, review-heavy, or release-mode presets.

3. Escalation thresholds

• Define exactly when the agent must ask before proceeding:
  • destructive changes
  • cost-bearing cloud calls
  • architecture changes
  • unclear scope

4. Proxy selection checklist

• Evaluate compatibility, streaming, tool history, hosted tools, auth, logging, and human-input behavior before choosing a proxy.

5. Conflict policy

• State that higher-priority system/developer rules override project files.
• State that direct current-turn user requests override project defaults.

Resources

scripts/

• sync_project_rules.py
  • Initialize a starter 3-profile bundle.
  • Render and sync AGENT.md / CLAUDE.md.
  • Preserve non-managed file content when possible.
• sync_response_style.py
  • Backward-compatible wrapper for style-only usage.

references/

• style-profile-schema.md
• agent-policy-schema.md
• llm-policy-schema.md
• proxy-selection-checklist.md

assets/

• response-style.profile.template.json
• agent-policy.profile.template.json
• llm-policy.profile.template.json
• AGENT.md.template
• CLAUDE.md.template

Minimal validation

After editing the profiles or templates:

• run python3 scripts/sync_project_rules.py --init-bundle-dir <tmp/rules> once
• run python3 scripts/sync_project_rules.py --project <tmp/project> --style-profile ... --agent-policy ... --llm-policy ...
• inspect generated AGENT.md and CLAUDE.md
• run the legacy wrapper once to confirm backward compatibility
• run ../../validate.sh from the package root before Git push