agenticnotetaking/skills/meta

/eidos:meta

Analyse the output of an eidos skill and drive a structured feedback loop to surface quality issues, abstraction-level problems, and improvement directions.

Usage

/eidos:meta [file or paste]      # analyse specific output
/eidos:meta                      # look for recent skill output in conversation

Instructions

1. Identify the Output

• File arg → read the file (could be a pull output, drift report, spec draft, etc.)
• Pasted content → use the pasted text directly
• No arg → scan the current conversation for the most recent skill output (look for patterns like <pull>, <spec>, drift reports, coherence findings)

Identify which skill produced the output (pull, drift, coherence, push, etc.) and what it was targeting.

2. Analyse

Evaluate the output against these dimensions:

• Abstraction level — is it at the right layer? (behaviour vs technique vs code detail)
  • Use the swappability test: could you swap the mechanism and the claim still holds?
  • Technicals that encode design preferences are fine; pure mechanism is not
• Completeness — does it cover what it should? Missing sections, gaps in coverage?
• Coherence — does it contradict itself or existing specs?
• Signal vs noise — is every sentence earning its place, or is there filler?
• Spec-worthiness — would you downstream code from this? Would it survive a rewrite of the implementation?

Not all dimensions apply to every output. Focus on what matters for the specific skill and context.

3. Structured Feedback

Present findings grouped by topic using letter/number notation:

Meta: [skill name] output on [target]

A. [First topic]
   Observation: [what you noticed]
   Example: [concrete quote or reference from the output]
   Question: [specific question for the human]

B. [Second topic]
   B1. [Sub-point]
       Observation: ...
   B2. [Sub-point]
       Observation: ...

Wait for the human's responses. Responses can be:

• Agreement or disagreement with reasoning
• A redirect ("actually the real issue is...")
• "skip" for topics to defer
• Elaboration that refines the observation

Multiple rounds are expected. Follow up on responses, go deeper where the human engages, drop threads they skip.

4. Create Meta Trace

When the dialogue reaches natural conclusions, run date '+%y%m%d%H%M' to get the current timestamp. Create memory/meta - <timestamp> - <claim>.md (per [[spec - naming - prefixes structure filenames as prefix claim pairs]], e.g. meta - 2602141030 - pull output abstraction level.md):

---
tldr: [Analysis of skill output — brief summary of what was found]
category: observation
---

# Meta: [what was analysed]

Context: [which skill, what target, what prompted the analysis]

---

## A. [Topic]

**Observation:** [what was noticed]
**Human:** [verbatim or close paraphrase of their response]
**Resolution:** [what this means — a concrete takeaway]

---

## B. [Topic]
...

---

## Next Steps

- [ ] [Concrete actions that emerged — spec updates, skill changes, claims to extract]

Show the draft to the user. If approved, write and commit.

5. Act on Outcomes (optional)

If the meta analysis surfaces concrete changes:

• Offer to update the skill or spec that produced the output
• Offer to extract claims (e.g. [[c - ...]])
• Offer to create todos for deferred items
• Offer to run the skill again with the improvements

Don't push — present options and let the human decide.

Output

• Creates: memory/meta - <timestamp> - <claim>.md
• May trigger: spec updates, skill updates, new claims, new todos