| name | recursive-self-improvement |
| description | Improve workflows for future agents. |
Improve workflows for future agents.
$ARGUMENTS
Process
Step 0: Update the specific file in .agents/skills/
Update the specific file in .agents/skills/ for the previous command that you were given, or for this specific override provided by the user.
Put in advice for future agents to collaborate better with the user, based on the user feedback that's been given to you in this interaction.
For example, if you were just asked to /implement-spec, then update .agents/skills/implement-spec/SKILL.md with this advice.
Important: The "previous command" is the slash command you were invoked with (look for <command-name>/command-name</command-name> in the conversation), NOT files you happened to edit while executing that command. If /debug-script-compilation led you to edit .agents/skills/compile/SKILL.md, you should still update .agents/skills/debug-script-compilation/SKILL.md.
Finding the previous command: Search backward from this /recursive-self-improvement invocation for the most recent <command-name>/...</command-name> tag. Don't assume "no command was run" just because the work felt like general debugging—slash commands often trigger investigative work.
Common mistake: If /systematize created document-implementation.md, the command to update is systematize.md, not document-implementation.md. Always search for <command-name> tags first.
If the feedback belongs in cross-cutting guidance rather than in a single process step, update the target script's ## Key Principles section. If the target script does not already have ## Key Principles, create that section only when you actually need it for the new guidance.
Step 1: Update compiled version if applicable
If the script has a compiled version (mekara.py inside the skill folder), update the compiled version too. Both must stay in sync.
Step 2: Commit the workflow improvement
Commit the workflow improvement after making the update.
Key Principles
- Don't edit this file unless it was invoked specifically for you to improve your own self-improvement
- Keep additions concise - only add what directly addresses user feedback
- Don't expand beyond the specific feedback given
- One line is often sufficient for simple guidance
- Add context about when the lesson applies
- Make it actionable with specific steps or checks to perform
- Include examples where helpful
- Integrate with existing sections instead of creating your own. This means editing an existing bullet point to expand it, NOT adding a new bullet point with related content. If there's already a guideline about "Document X", add your insight as a sub-bullet or expand the text inline—don't create a separate "Document Y" bullet.
- Keep it generalizable. Don't say "Always have ClassB subclass ClassA"; say instead "All Controller classes should subclass BaseController." If you include the specific example of "ClassB subclasses ClassA," make it clear how that example generalizes.