Reflect
Mine the current conversation for durable learnings, then route them into skill edits.
When to invoke
- The user said "reflect" or "/reflect".
- A complex task (5+ tool calls) just landed cleanly and the recipe is worth keeping.
- The agent hit dead ends, found the working path, and the path generalizes.
- The user corrected the agent's approach mid-task.
- A non-trivial workflow emerged that isn't captured anywhere.
Skip when the conversation is trivial, off-topic, or already covered by an existing skill the parent followed correctly. One-offs are not learnings.
Process
1. Locate the active transcript
The parent finds its own transcript file before fanning out. Transcript location is harness-specific — Claude Code keeps them under ~/.claude/projects/<slug>/, Cursor under ~/.cursor/projects/<slug>/agent-transcripts/. If the system prompt names a transcript directory for the active workspace, that path wins. Scope the glob to the active workspace's directory only. Globbing across ~/.<harness>/projects/*/ crosses workspace boundaries and reads private chats from unrelated projects.
ls -t <transcript-dir>/*.jsonl <transcript-dir>/*/*.jsonl <transcript-dir>/*/subagents/*.jsonl 2>/dev/null | head -10
Three transcript layouts to expect: flat (<id>.jsonl), nested (<id>/<id>.jsonl), and subagent (<parent>/subagents/<child>.jsonl).
For each candidate, read the first JSONL line and check that message.content[0].text contains the conversation's opening user prompt. Take the matching path. If no path resolves, write a tight digest of the session and pass that instead.
2. Spawn three reviewers in parallel
One message, three general-purpose subagents. Give them tool access rather than read-only mode: reviewers need MCP access for context lookups (tickets, chat threads, observability traces referenced in the transcript), and read-only modes commonly strip MCPs. The prompt forbids file writes; the parent applies edits.
| Lens | Prompt template |
|---|---|
| Judgment | references/judgment-reviewer.md |
| Tooling | references/tooling-reviewer.md |
| Divergent | references/divergent-reviewer.md |
Pass each template verbatim, substituting the transcript path or digest where marked. Reviewers return findings in their response bodies.
If the harness lets you set a model or reasoning effort per subagent, give the divergent lens a different model from the judgment lens — the three lenses exist to disagree, and identical runners weaken that. Where it does not, the distinct prompts still carry most of the separation.
3. Synthesize
One general-purpose subagent, with tool access rather than read-only: the synthesizer's quality check includes spot-verifying citations, which can require MCP access. Use references/synthesizer.md verbatim, with each reviewer's full output inlined where marked. The synthesizer returns a structured Accepted / Rejected / Backlog list.
4. Structural enforcement check
Sanity-check the synthesizer's Accepted list. For any item that would be enforced more reliably by a lint rule, script, metadata flag, or runtime check, move it from Accepted to Backlog. The synthesizer already applies this criterion; this is a final pass before edits land. See the encode-lessons-in-structure principle.
5. Apply
Before applying any Accepted edit, present the synthesizer's full Accepted/Rejected/Backlog output to the user and wait for explicit approval. The user picks which subset to apply and may redirect routings. Skill changes affect every future agent in the org; do not auto-apply.
Backlog items file to whatever devex / backlog tracker your team uses automatically. Those are tracker submissions, not skill edits. Only the Accepted list waits for approval.
For each approved Accepted item, follow the Routing field exactly:
- Trivial existing-skill edit (a one-line bullet, a tightened sentence, a stale fact corrected): parent does directly.
- Substantive existing-skill edit (a new section, a new pattern table, more than ~10 lines): hand to whatever skill-authoring skill this harness ships (Claude Code has
skill-creator; several harnesses ship acreate-skill) and run its draft / test / iterate loop. If none exists, draft the edit, then re-read the whole skill cold and cut anything the edit made redundant. tune description: <skill path>(the skill exists but didn't trigger when it should have): the description is the entire trigger mechanism. Rewrite it to lead with the use case and name the concrete words a user would actually say. Hand to the authoring skill's description-optimization loop where one exists.new skill: <kebab-name>: hand creation to the authoring skill. Do not invent the shape ad hoc — a new skill needs aname, a trigger-shapeddescription, and bulk content inreferences/rather than in the body.
If your environment ships a SKILL.md validator, run it on every touched skill before declaring done. Skip this step if it doesn't.
6. Summarize for the user
Short list, no preamble:
- Edits applied:
<skill path>. What changed, one line each. - New skills created:
<skill path>. One line each (rare). - Backlog filed to the devex tracker:
<issue title>(<tags>). One line each. - Dropped: one line per rejected finding + reason from the synthesizer.