Improving agent reflection

Offer opportunities to continuously improve agent rules and skills, based on user interactions and feedback.

Sequential steps to be followed when using this skill

When improving agent reflection, follow those steps.

Create the improving-agent-reflection Execution Checklist (MANDATORY)

• Before executing anything, create a checklist named improving-agent-reflection Execution Checklist with all steps of these instructions.
• The improving-agent-reflection Execution Checklist must include all numbered steps explicitly.
• After completing each step of these instructions, mark the item in the improving-agent-reflection Execution Checklist as completed.
• Do not skip any item.
• If an item cannot be executed, explicitly explain why.
• Never mark the task as completed while any item from the improving-agent-reflection Execution Checklist remains open.

1. Inform the user

• Always tell the user "SKILL: I am improving agent reflection" to inform the user that you are running this skill.

2. Applicability Check

• Always use agent: ask_followup_question to ask whether reflection is warranted for this task:

  <ask_followup_question>
  <question>Did this task involve user feedback at any point OR multiple non-trivial steps (e.g., several file edits, complex logic generation)?</question>
  <options>["Yes - proceed", "No - skip"]</options>
  </ask_followup_question>

• If "No - skip", conclude briefly and stop this skill.

3. Offer Reflection

• Always use agent: ask_followup_question to confirm the user wants reflection and proposals:

  <ask_followup_question>
  <question>Before I proceed, would you like me to reflect on our interaction and suggest potential improvements to the active rules and skills?</question>
  <options>["Yes - reflect", "No - skip"]</options>
  </ask_followup_question>

• If "No - skip", conclude briefly and stop this skill.

4. Identify Active Rules

• Always use the agent: read_file tool to read the file AGENTS.md, and all the referenced *..md files referenced from it. This file contains rules that are always active.

  <read_file>
  <path>AGENTS.md</path>
  </read_file>

5. Identify Active Skills (Best‑Effort)

• Skills directory: .cline/skills
• Always use the agent: list_files tool to list all skills that are present:

  <list_files>
  <path>.cline/skills</path>
  </list_files>

6. Load Accessible Skill Files

• For each accessible skill that you want to consider, always use the agent: read_file tool to read the SKILL.md file of the skill, and all the referenced *..md files referenced from it.

  <read_file>
  <path>.cline/skills/{skill-name}/SKILL.md</path>
  </read_file>

• If the content is referencing other files (*.md), always use the agent: read_file tool to read them:

  <read_file>
  <path>.cline/skills/{skill-name}/references/{referenced-file}.md</path>
  </read_file>

7. Review and Synthesize Opportunities

Using the current task's conversation context:

• Always summarize relevant user feedback (explicit and implicit)
• Always identify where rules and skills helped or hindered flow
• Always propose targeted improvements focused on:
  • Addressing user feedback directly
  • Improving clarity and conciseness
  • Consolidating overlapping guidance
  • Removing outdated or low-impact sections

8. Present Proposals

• Always provide a concise list of suggested changes per file (no diff blocks). Keep focus on practical, high-value edits.

9. Approval to Act

• Always use agent: ask_followup_question to ask the user how to proceed:

  <ask_followup_question>
  <question>Would you like me to apply these improvements now where possible, or just present recommendations?</question>
  <options>["Apply now", "Show recommendations only", "Cancel"]</options>
  </ask_followup_question>

10. Execute or Report

• If "Apply now": implement updates for accessible files
• If "Show recommendations only": present a clean summary for manual application
• If "Cancel": make no changes

11. Conclude

• Always summarize what changed or what to change next, then finish.

Final Verification (MANDATORY)

Before declaring the task complete:

• Re-list all numbered steps from the improving-agent-reflection Execution Checklist.
• Confirm each one was executed.
• If any step was not executed, execute it now.

Constraint

Do not offer reflection if:

• No rules or skills were active.
• The task was very simple and involved no feedback.

When to use it

• Always use it every time another skill specifically mentions skill: improving-agent-reflection.
• Always use it every time the user asks you to improve rules or skills.
• Always use it before attempting completion of a task that involved user feedback on active skills and rules.