Tune memory and planning instructions for agents#658
Conversation
…irectory Replace bulk-load-everything approach with a targeted pointer: read .agents/quick-reference-card.md first, discover skills from .agents/skills/, and consult .agents/_TOC.md only when a skill does not cover the context. Incorporate workflow rules, self-improvement loop, verification and quality guidelines, core principles, and task flow from the proposed CLAUDE-02 template. Adjust task paths to .agents/tasks/ so todo.md and lessons.md are shared across all AI agents. Add .agents/tasks/todo.md and .agents/tasks/lessons.md stubs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f79810b190
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| @@ -0,0 +1,8 @@ | |||
| { | |||
There was a problem hiding this comment.
Because migrate copies the whole .claude directory into every consuming repo, committing .claude/settings.local.json propagates a file that Claude Code treats as per-user/local and normally gitignored rather than shared project config (Claude docs). If the pre-PR skill permission is meant for all collaborators, it should live in .claude/settings.json; otherwise this file should be ignored so local permission overrides do not get distributed or overwritten on every config update.
Useful? React with 👍 / 👎.
Claude system prompt says the opposite: "Subagents are valuable for parallelizing independent queries or for protecting the main context window from excessive results, but they should not be used excessively when not needed."
alexander-yevsyukov
changed the title
2 participants
This PR improves the instructions for agents encouraging updating memory and structuring working with plans. The primary focus is made on Claude Code, but instructions are compatible with other AI agents too.
Updating instructions were guided by CC so they should be efficient with the current version of CC and its system prompts.