| name | reflect |
| description | Analyzes current conversation for user corrections, patterns, and preferences, then proposes skill improvements. Use after sessions with corrections or when patterns emerge from successful approaches.
|
| disable-model-invocation | true |
| allowed-tools | Read, Edit, Glob, Grep, AskUserQuestion |
Reflect
Analyze the current conversation to find signals for improving existing skills.
This skill handles patch-level remediation only — fixing steps, criteria, format,
or behavior of a skill that already exists. If the problem requires a new skill,
a subagent, a hook, MCP integration, or a CLAUDE.md rule, use /reflect:reflect-architecture
instead.
Steps
1. Scan conversation for signals
Look for:
- Corrections (HIGH) — "Don't do X, do Y instead", user rejecting an approach
- Approvals (MEDIUM) — "perfect", "exactly right", confirming a non-obvious pattern works
- Preferences — "always use...", "never do...", explicit workflow choices
- Bugs in skills — skill gave wrong command, wrong URL, wrong API format
Ignore:
- Teammate/agent messages (not user corrections)
- Quoted text (user referencing, not correcting)
- Generic praise without a specific reusable pattern
- One-off instructions specific only to this task
2. Filter to skill-level fixes
For each signal, check:
- Does an existing skill already cover this area?
- Is the fix a change to that skill's steps, criteria, or output format?
If the answer to either is "no", note it as an architecture signal and suggest
running /reflect:reflect-architecture after this skill completes. Do not attempt
to create new skills, hooks, or CLAUDE.md rules from here.
3. Present findings
For each signal that maps to an existing skill, show:
- Which skill it affects (name and file path)
- What the signal is (quote the user)
- Confidence: HIGH / MEDIUM / LOW
- Proposed change to the skill
If nothing found — say so and stop.
4. Get approval
Ask the user:
- Apply all
- Review individually
- Skip
5. Apply
Edit the target SKILL.md files directly with the approved changes.
After editing, remind the user: if an edited skill was already invoked in this
session, it must be re-invoked for the changes to take effect.
