AgentSkillsCN

refinement-loop

通过多次迭代逐步优化。当用户要求“深入思考”、“提炼”、“精炼”或“反复迭代”某件事时，或当问题更适合多次尝试而非一次完成时，可采用此方法。

SKILL.md
--- frontmatter
name: refinement-loop
description: Iterative refinement through multiple passes. Use when the user asks to 'meditate on', 'distill', 'refine', or 'iterate on' something, or proactively when a problem benefits from multiple passes rather than a single attempt.

Refinement Loop

STARTER_CHARACTER = 🔄

Iterative refinement through file artifacts. Each pass removes one layer of noise, revealing the next.

Setup

Ensure playground/ exists and is in .gitignore. All iteration files go there.

Process

1. Clarify Goal (if the user hasn't defined it for you already earlier in the conversation)

Ask the user:

  • •What are you refining and what is the goal of refinement?
  • •Derive a short tag that we'll use as filename for iterating: {goal}-{subject} (e.g., gist-nullables, simplify-api, distill-auth-docs)

2. Capture Starting Point

Write original to: playground/{goal}-{subject}-0.md

3. Iterate

Loop:

  1. •Read back the current file (forces fresh perspective)
  2. •Reflect critically: What's missing? What's weak? What could be clearer?
  3. •If improvements found: Write improved version to playground/{goal}-{subject}-{N+1}.md, then loop again

Before Stopping - Exhaustive Check

When you think you're done, you're probably not. Run through this:

  1. •List everything that could still be improved - even small things (formatting, word choice, structure, clarity)
  2. •Consider what we haven't considered - what angles did we miss? What would someone else notice?
  3. •Try improving in a new direction - not just polishing what's there, but questioning assumptions
  4. •Read as if seeing it for the first time - does it immediately make sense? Is anything unclear?

Only stop when you've gone through this checklist extensively multiple times and genuinely found nothing. There is no "good enough" - someone will use this later and shouldn't waste time on mediocre results.

4. Present Final

Show the user the final version with a brief summary of the refinement journey and number of iterations you used. If deeper issues or questions surfaced, present them to the user as well.

Principles

  • •Read back forces fresh eyes: Reading from file breaks the "I just wrote this" blindness
  • •No "good enough": Every detail matters. Formatting, word choice, structure - all of it
  • •Consider the unconsidered: What angles haven't we explored? What would someone else see? Is there something that the user hasn't even considered, too? (you can surface it as questions at the end)
  • •Files force iterative improvement: Writing to files prevents "pretend" iteration in conversation. Real iteration gets to much better results.
  • •Earn the stop: Run the exhaustive check. If you find nothing, you've earned stopping.