Skill reflect

Skill Reflect

You run immediately after a skill session ends. Your job is to mine the conversation that just happened, extract the friction points, and write a structured log to disk so the evolver can read it later. The session is still in context — use it.

---

Phase 1: Resolve the skill

The user invokes this as /skill-reflect <skill-name> (e.g., /skill-reflect qa). If they didn't provide a skill name, ask: "Which skill should I write a session log for?"

Once you have the name, find its owning plugin. Try scanning the working directory first (works if you're in the skills repo):

ls -d plugins/*/skills/<skill-name>/ 2>/dev/null

• One match — record the plugin name (the second path segment, e.g. plugins/<plugin>/skills/...).
• Multiple matches — ask the user to qualify: "I found <skill-name> in both <plugin-a> and <plugin-b> — which one? (e.g., <plugin-a>/skill-name)"
• No match — before asking the user, check whether you're in the skills repo on the wrong branch: run git branch --show-current. If plugins/ exists but the scan found nothing, the skill's plugin may live on an unmerged branch — switch to the correct branch and re-scan. Only if the branch check doesn't resolve it, ask: "Which plugin does <skill-name> belong to? (delivery, writing, track, or meta)" and use their answer as the plugin name.

---

Phase 2: Reflect on the session

Review the current conversation — the skill run is still in context. Work through each of the four questions below. Be specific: quote or paraphrase actual moments from the conversation rather than speaking in generalities. A vague log is useless to the evolver.

2.1 What did you have to improvise?

Look for steps you added on the fly that weren't instructed by the SKILL.md — decisions you made without a written rule to follow, phases you invented, tools you reached for that the skill didn't mention. Improvisation is a signal the SKILL.md has a gap.

Hint: scan for moments where you reasoned "I'll also…", "it makes sense to…", or added a step that wasn't in a numbered phase.

2.2 What did the user correct?

Look for explicit corrections, redirections, and pushback: "no", "wait", "actually", "that's not right", "don't do that", or any message where the user steered you away from what you were doing. Each correction is a place where the skill's instructions were unclear, missing, or wrong.

2.3 What edge cases came up that the skill didn't anticipate?

Look for situations you had to navigate that the SKILL.md didn't prepare you for — ambiguous inputs, missing context, unusual project state, dependencies the skill assumed existed but didn't, questions you had to ask that a better-written skill would have answered upfront.

2.4 What specific changes would help?

Based on the above, propose concrete improvements to the SKILL.md. Be specific about what would change: a new step in a phase, a clarification to an existing instruction, a fallback for an edge case, a new question to ask upfront. Vague suggestions ("improve the instructions") are not useful — name the section and describe the change.

---

Phase 3: Write the log

Get the current timestamp:

date -u +"%Y-%m-%dT%H-%M"

Create the directory and write the log:

mkdir -p ~/.claude/skill-sessions/<plugin>/sessions/<skill-name>

Write to ~/.claude/skill-sessions/<plugin>/sessions/<skill-name>/<timestamp>.md using this exact structure:

---
skill: <skill-name>
plugin: <plugin>
date: <YYYY-MM-DD>
session_summary: <one sentence describing what was done in this session>
---

## What I had to improvise

<your findings from 2.1, or "Nothing notable." if clean>

## User corrections

<your findings from 2.2, or "No corrections." if clean>

## Edge cases not anticipated

<your findings from 2.3, or "None encountered." if clean>

## Proposed improvements

<your findings from 2.4, or "No changes recommended." if clean>

If all four sections are empty (a clean run with no friction), still write the log — a record of clean sessions is signal too.

---

Phase 4: Confirm

Tell the user:

1. The full path of the file written.
2. The single most actionable friction point from this session, in one sentence. If the session was clean, say so.

Example:

Session log written to ~/.claude/skill-sessions/delivery/sessions/qa/2026-06-04T14-32.md. Top friction: the skill didn't handle the case where no Linear issue could be inferred from the branch name — Claude had to ask, and a fallback prompt should be in Phase 0.

---

What to skip

• Don't modify the SKILL.md itself — that's the evolver's job, not yours.
• Don't summarize the whole session — focus on friction, not narrative.
• Don't be diplomatic. If the skill's instructions were unclear, say so plainly.