harvest-automation

name	harvest-automation
description	Use when the user wants to turn past Claude Code work into reusable automation — invoked as /harvest-automation, optionally with a window like "7d". Triggers include end-of-session review, "what should we automate", "mine my recent sessions", recurring manual workflows, repeated corrections, repeated permission prompts, or preferences worth capturing as a skill, CLAUDE.md entry, or memory.
disable-model-invocation	true

Harvest Automation

Mine past Claude Code work for recurring patterns and turn them into durable, reusable automation: skills, CLAUDE.md entries, memory, and permission allowlists.

Local skill, maintained in this repo (.source.json has "repo": null). Replaces the old retrospect and workflow-from-chats skills.

Core principle

Capture a pattern only if it recurs (≥ 2 occurrences across the scope) or the user explicitly asks to keep it; one-offs are noise. Every claim needs evidence: session id, turn number, a verbatim quote (≤ 15 words), or a tool-call signature.

When to use

Runs only when invoked explicitly: /harvest-automation analyzes the current session; /harvest-automation 7d (or "last week") analyzes a window of recent sessions. Decline when the user asked for a plain summary, the scope has < ~4 user-turns of real interaction, or there were no tool calls.

Scope and gathering

Default scope is the current session. Reflect on the in-context conversation first.

• For a window (Nd), run {baseDir}/find-sessions.sh --since N to list recent session JSONLs across all project dirs in the active profile, newest first. Recurring automation often spans repos, so scan across projects, not just the current one. Read enough of each to extract signals; cite the session by id/path.

Distill — recurrence-first signal detection

Cluster evidence into candidate patterns. For each, record: trigger, evidence, recurrence count, confidence, and the artifact it should become.

Recurring pattern	Becomes
A multi-step tool sequence done manually ≥ 2× (e.g. fetch→categorize→render)	Skill
The same correction / nudge (no, don't, instead, that's not…)	CLAUDE.md (project fact) or memory (preference)
A stated preference repeated (I prefer, always, never)	memory
The same tool + command approved repeatedly	permission allowlist
Re-explaining the same project context across turns/sessions	CLAUDE.md
Repeated web lookups for the same library docs	memory note / use context7

Also flag wasted-effort smells as supporting evidence (not artifacts on their own): same file Read ≥ 2× with overlapping ranges, broad search followed by a narrow one on the same corpus, a subagent dispatched for a one-grep task, retries of a failing command without diagnosis between attempts.

Confidence

• Strong — explicit user preference, a workflow-changing correction, or the same pattern across ≥ 2 sessions.
• Medium — an accepted workflow or a repeated tool/validation choice the user relied on.
• Weak — agent-chosen behavior with no user feedback, or a single ambiguous instance.
• Contradicted — evidence points both ways; ask the user before writing anything.

Promote only Strong/Medium to artifacts. List Weak as "consider"; never auto-apply it.

Materialize — artifact routing

Route each promoted pattern to exactly one artifact. Stay composable — delegate, do not reimplement.

• Skill — a recurring multi-step workflow with clear triggers. Draft a proposal (name, "Use when…" description, trigger, the steps, any scripts), then hand off to superpowers:writing-skills to author and validate it under skills/<name>/; mark it local with a {"repo": null} .source.json. Never scaffold skill files by hand.
• CLAUDE.md — a collaborator-visible project fact (build/test commands, invariants, code locations, "always use X helper"). Applied via {baseDir}/apply-suggestions.sh.
• Memory — a user-private preference (terse vs verbose, tool choices, workflow habits). Applied via {baseDir}/apply-suggestions.sh.
• Permission allowlist — repeated approvals of the same tool/command. Reference the fewer-permission-prompts skill to generate the .claude/settings.json entries; do not re-implement its logic.

CLAUDE.md vs. memory

CLAUDE.md is collaborator-visible; memory is user-private.

• CLAUDE.md-worthy — project facts any contributor benefits from. Before proposing, Read the project-root CLAUDE.md (via git rev-parse --show-toplevel) so proposals don't duplicate it.
• memory-worthy — anything phrased as "I prefer" / "I always"; personal workflow habits.
• neither — one-off events. Require ≥ 2 occurrences OR explicit user feedback ("remember this") to promote; otherwise drop.

Report shape

Emit exactly this structure, organized by artifact type. Keep it under ~400 lines; compress, don't truncate.

# Harvest Automation

## Scope
<current session | last N days: M sessions across K projects>
- Interaction: <N user / M assistant turns>

## Summary
<P patterns found> → <A skills, B CLAUDE.md, C memory, D permission>

## Proposed skills
- <name> — <"use when" trigger>. Recurs <N>× (<evidence>). → hand to writing-skills.

## Proposed CLAUDE.md additions
Target file: <absolute path, or "propose new at <path>">
(Unified diff — exactly what apply will write.)

## Proposed memory entries
- name: <slug>
  type: <feedback|user|project|reference>
  description: <one-liner>
  body: |
    <content; Why: and How to apply: for feedback/project types>

## Proposed permission allowlist
- <tool + command pattern> approved <N>× → see fewer-permission-prompts.

## Consider (weak signal, not auto-applied)
- <pattern> — <evidence>.

## Apply?
Reply: `apply skills`, `apply claude`, `apply memory`, `apply all`, or `skip`.

Apply on confirmation

• apply skills — for each proposed skill, invoke superpowers:writing-skills with the drafted proposal. Do not write skill files directly.
• apply claude / apply memory / apply all — Write the payload (schema below) to /tmp/harvest-automation-<sessionid-or-timestamp>.json, then call:

{baseDir}/apply-suggestions.sh <claude|memory|all> <payload.json>

{
  "claude_md": { "path": "/abs/CLAUDE.md", "append": "## Section\n\nBody…\n" },
  "memory": {
    "dir": "/abs/memory",
    "entries": [
      { "filename": "feedback_terse.md",
        "content": "---\nname: …\ndescription: …\nmetadata:\n  type: feedback\n---\n\n<body>\n",
        "index_line": "- [Terse output](feedback_terse.md) — prefers short answers" }
    ]
  }
}

(apply all runs CLAUDE.md + memory; skills are still applied separately via writing-skills.)

Anti-patterns

• Do NOT dump the transcript. The report is the analysis, not a log.
• Do NOT edit CLAUDE.md or memory files directly — always go through apply-suggestions.sh so backups and index updates stay consistent.
• Do NOT re-run tools that already ran in-session to re-verify. Reuse prior results.
• Do NOT use performative praise or emojis. State a genuine strength in one line under ## Strengths only if there is one; otherwise omit.

Scripts

{baseDir}/find-sessions.sh                 # current session JSONL (most recent for cwd)
{baseDir}/find-sessions.sh --since 7       # recent session JSONLs (last 7 days), newest first
{baseDir}/apply-suggestions.sh <scope> <payload.json>   # apply CLAUDE.md / memory edits

Troubleshooting

If context was compacted (system-reminders mention compaction, or the earliest turns are missing), run {baseDir}/find-sessions.sh to locate the session JSONL and Read the earliest portion to recover missed turns.