jayden-dang Reflect
Turn a finished hard session into skill-set improvements. Do not auto-apply.
When
User said /reflect, or a complex task (many tool calls) just landed with a
generalizable path, dead-ends, or mid-task corrections. Skip trivial chats.
Steps
- Locate transcript. Prefer the harness transcript path for this workspace only. Do not glob other projects' private chats. If none, write a tight digest of the session.
- Three lenses in parallel (read-only subagents; different models when
possible): Judgment (what should have been non-negotiable), Tooling (what
script/lint/check would have caught it), Divergent (what a skeptical peer
would refuse). Templates in
references/beside this file. - Synthesize into Accepted / Rejected / Backlog. Prefer encoding lessons
as lint, test, metadata, or schema over more prose
(
encode-lessons-in-structure). - Present and wait. Show the full list. Apply only what the user approves.
- Apply via author-skills. Trivial one-line fixes: parent edits.
Substantive edits or new skills: name
/author-skillsand follow its RED→GREEN cycle — reflect never skips the failing-test bar for skill text. - Summarize edits applied, new skills, backlog filed, drops with reasons.
Rationalizations
| Thought | Reality |
|---|---|
| "I'll just patch the skill — faster" | Unapproved skill edits hit every future agent. Wait |
| "author-skills is enough; skip reflect" | OVERRIDE: user wanted this entry point |
| "Backlog is optional" | File it; Accepted waits for approval |
Done when
Approved Accepted items are applied under /author-skills discipline; Rejected
and Backlog are stated; no silent skill writes.