Refine

/refine — continual harness refinement

Port of prime-agent's Continual Harness for Claude Code. Reviews what happened in the current session and persists a small number of durable lessons into the memory system, each backed by quoted evidence, each snapshotted for rollback.

Invocations

• /refine — review the trajectory, propose and apply refinements
• /refine <instructions> — refine with focus (e.g. "only capture the deploy workflow")
• /refine rollback <id> — undo a recorded refinement
• /refine history — list recent refinements from the log

Storage layout

• Harness state = the persistent memory directory (listed in your system prompt's Memory section) — memory files + MEMORY.md index. That is the supplemental state this skill edits.
• Refinement log = ~/.claude/harness/refinements.jsonl (create the directory if missing). One JSON object per line.

Refinement procedure

1. Review the trajectory. Scan the conversation for durable lessons: corrections the user made, approaches that worked after failures, workflow patterns repeated 2+ times, project constraints discovered the hard way, reusable subagent/tool recipes. Ignore anything session-specific, already-recorded, or derivable from the repo itself.

2. Propose at most 5 edits. Each edit is one of:

   • create — a new memory file (kinds: user, feedback, project, reference)
   • update — revise an existing memory file that this session proved wrong or incomplete
   • delete — remove a memory this session invalidated

   Every edit MUST carry evidence: a short verbatim quote (or precise paraphrase with location) from THIS conversation showing why the edit is warranted. No evidence, no edit. Prefer fewer, higher-confidence edits.

3. Hard rules (immutability):

   • NEVER edit ~/.claude/CLAUDE.md, project CLAUDE.md, or any settings file — those are the immutable base prompt, exactly as in prime-agent.
   • Only memory files and MEMORY.md index lines may change.
   • Respect the existing memory frontmatter format; keep entries short.

4. Apply and log. For each applied edit, append one line to ~/.claude/harness/refinements.jsonl:

   {"id": "<8-hex>", "ts": "<ISO-8601>", "action": "create|update|delete",
    "file": "<memory file path>", "evidence": "<the quote>",
    "before": <full prior file content as string, or null>,
    "after": <full new file content as string, or null>,
    "index_before": "<prior MEMORY.md line or null>",
    "index_after": "<new MEMORY.md line or null>"}
   

   Generate the id with 8 random hex chars. before/after snapshots are the rollback mechanism — never skip them.

5. Report. Tell the user what was created/updated/deleted, one line each with its refinement id and the evidence, so rollback ids are visible.

Rollback procedure (/refine rollback <id>)

1. Find the line with matching id in ~/.claude/harness/refinements.jsonl.
2. Invert it: create → delete the file and its index line; update → restore before content and index_before; delete → recreate the file from before and restore its index line.
3. Append a new log line recording the rollback (action rollback, with the inverse snapshots), and confirm to the user.

Mapping prime-agent kinds → memory types

• prompt note → feedback memory (how to behave, with Why and How to apply)
• memory → user or project memory (facts, preferences, constraints)
• skill description → reference memory describing a reusable call, e.g. an rlm-repl py_exec recipe or a shell one-liner, with exact invocation
• subagent spec → reference memory containing a reusable Agent-tool prompt (agent type, prompt template, when to use)

Calibration

Prime-agent's own weakness (verified by source audit): "evidence-backed" was prompt-only there. Here the evidence quote is mandatory in every log line, and an edit without a verifiable quote from the current conversation must be dropped. When in doubt, propose nothing — an empty refinement is a valid outcome and better than memory pollution.