golden-evolver

name	golden-evolver
description	Evolve the golden few-shot examples your brain's processes use. The owner names a task + a recipe; the skill runs the evolution cycle (present outputs, collect the owner's scores, track versions, propose mutations) and saves a versioned golden set the other skills read at prompt-build time. Tasks: classification (taxonomy routing, used by /ingest-brain-dump + /mine), summarization (Template A/B summaries, used by /process-content), keyword_extraction (used by /process-content), template_creation (full Template A from raw input), segmentation (idea-boundary detection in long text), and delegation (orchestrator → sub-agent prompts; see Orientation_Docs/SUBAGENT_DELEGATION_PRIOR_ART.md). Recipes: Quick Curation (hand-pick 5), Bootstrap (generate 20, filter to top 5), Deep Evolution (meta-prompt iteration), Escape Local Maximum (aggressive mutation + wildcards), Hybrid (bootstrap then evolve). Invoke: "/golden-evolver", "/golden-evolver classification", "evolve examples for keyword extraction", "quick curation for summarization", "escape local max on segmentation".
required_context_files	["Orientation_Docs/CONTENT_TAXONOMY.md"]
user_invocable	true
allowed_tools	["Read","Write","Edit","Bash","Glob","Grep","Agent"]

Golden Few-Shot Evolver

Evolve the golden few-shot examples your brain's processes use. Better examples make every downstream run better — and they're cheap to improve once you have a loop. The owner names a task and a recipe; you manage the cycle: present outputs, collect scores, track versions, propose mutations.

Storage (self-contained — no external project needed)

.claude/golden_sets/[task]/active.json          # current golden set (the one skills read)
.claude/golden_sets/[task]/v001_[hash].json     # versioned sets
.claude/golden_sets/[task]/score_history.jsonl  # one line per scored generation

Create .claude/golden_sets/ on first use. It's local working state — add it to .gitignore if you don't want golden sets in version control.

Invocation

• /golden-evolver — interactive (ask which task)
• /golden-evolver classification — target a task
• /golden-evolver bootstrap summarization — recipe + task
• "evolve examples for keyword extraction" / "quick curation for classification" / "escape local max on segmentation"

Step 1 — Identify the task

Task	What it does	Where the examples are used
classification	Categorize content by taxonomy	/ingest-brain-dump, /mine
summarization	Generate Template A/B summaries	/process-content
keyword_extraction	Pull keywords from content	/process-content
template_creation	Build a full Template A from raw input	/process-content
segmentation	Detect idea boundaries in long text	any long-text split
delegation	Write orchestrator → sub-agent prompts	every Task/Agent dispatch (see Orientation_Docs/SUBAGENT_DELEGATION_PRIOR_ART.md)

If the owner doesn't specify, list the tasks and ask.

Step 2 — Load current state

ls .claude/golden_sets/[task]/. If active.json exists, read it + score_history.jsonl. If not, this is first-time setup — default to Quick Curation.

Step 3 — Pick a recipe

• "quick curation" / "let me pick" → Quick Curation
• "bootstrap" / "generate and filter" → Bootstrap
• "deep evolution" / "evolve" / "iterate" → Deep Evolution (meta-prompt)
• "escape local max" / "shake it up" → Escape Local Maximum
• "hybrid" → Hybrid (bootstrap then evolve)

No recipe given? Recommend by state: no set → Quick Curation; <5 scored generations → Bootstrap; 5+ → Deep Evolution.

Step 4 — Execute

Quick Curation — find 5 real inputs (grep your brain for relevant content) → run the task on each → present all 5 outputs → owner approves/rejects each → save the approved ones as v001_[hash].json + point active.json at it.

Bootstrap — find 20 real inputs → run the task on all (parallel cheap-model agents are fine) → present in batches of 5 → owner scores approve/reject → keep the top ~5 → save versioned + log all scores.

Deep Evolution — read active.json + score_history.jsonl → build a meta-prompt ("here are past sets and their scores; the current best scored X; propose 3 candidates that might score higher; include one wildcard to escape local maxima") → run the task on a test input with each candidate → present all 3 + the current best → owner ranks/scores → log scores, update active.json if a candidate wins → repeat until "good enough."

Escape Local Maximum — Deep Evolution with aggressive mutation, ~50% wildcard injection, crossover between top sets enabled.

Hybrid — Bootstrap to seed, then Deep Evolution to refine.

Step 5 — Save & report

Golden Evolver — [Task]
  Recipe: [...]   Version: v[N]_[hash]   Examples: [count]   Generations scored: [count]
  Best score this session: [score]
  Active set: .claude/golden_sets/[task]/active.json
  Iterate: "deep evolve [task]"

Step 6 — Integration

Other skills read .claude/golden_sets/[task]/active.json at prompt-build time and include those examples in the prompt; if it doesn't exist, they fall back to zero-shot. (The shipped skills work zero-shot today — golden sets are a quality upgrade you grow, not a dependency.)

Schemas

score_history.jsonl line: {"task","version","input_hash","score","score_type":"1-5","timestamp","notes"}

v[N]_[hash].json: {"task","version","hash","created","parent_version","strategy","examples":[{"input","output","source"}],"avg_score","generations_tested"}