autorefine

分类: 调研与分析 | 上传者: surahli123 | 下载: 0 | 版本: v1.0（最新）
Iterate and improve any skill using eval-grounded autoresearch. Combines v2.0 design audit, Hamel's Three Gulfs eval methodology, and Karpathy-style mutation optimization. Use when you want to assess skill quality, build evals from scratch, run error analysis, or optimize a skill through experiments.
更新日志: Source: GitHub https://github.com/surahli123/autorefine-skill-improvement
目录结构

当前层级: autorefine/
📁 assets/
- 📄 dashboard-ds-trace.png 198.7 KB
📄 dashboard.html 12.0 KB
📄 README.md 4.4 KB
📄 references.md 3.8 KB
📄 SKILL.md 17.6 KB
SKILL.md

---
name: autorefine
description: Iterate and improve any skill using eval-grounded autoresearch. Combines v2.0 design audit, Hamel's Three Gulfs eval methodology, and Karpathy-style mutation optimization. Use when you want to assess skill quality, build evals from scratch, run error analysis, or optimize a skill through experiments.
---

# AutoRefine

Guided pipeline for improving any skill — from zero evals to optimized and validated.

Three gulfs to cross, in order: **Comprehension** (what does this skill actually do?) → **Specification** (do our judges measure what matters?) → **Generalization** (does it work on unseen inputs?).

## Quick Start

Point at a skill directory: `/autorefine path/to/my-skill/`
The agent detects your progress and picks up where you left off.

## Preflight

1. **Target exists.** Confirm the path contains a SKILL.md. If not, ask for the correct path.
2. **Detect enhancements.** Search for Hamel's `eval-audit` and `error-analysis` skills. If found, note in state.json. These enhance but are NOT required.
3. **Report tier:** Full (Hamel's detected) or Basic (core methodology only, works on any agent with Read/Write/Bash).

## Initialize Workspace

If `autoresearch-<skill>/` doesn't exist in the target directory:

1. Create `autoresearch-<skill>/` and `autoresearch-<skill>/traces/`
2. Generate `state.json`:
```json
{"skill_name":"<name>","skill_path":"<path>","started":"<today>","current_phase":1,"current_gulf":1,"phases":{},"gates":{"gulf_1":"pending","gulf_2":"pending"},"hamel_available":false}
```
3. Generate empty `results.json`:
```json
{"skill_name":"<name>","status":"running","current_experiment":0,"baseline_score":null,"best_score":null,"experiments":[],"eval_breakdown":[]}
```
4. Generate `results.tsv` with header: `experiment\tscore\tmax_score\tpass_rate\tstatus\tdescription`
5. Generate empty `changelog.md` and `eval-suite.md` from formats shown in Phase 7
6. Copy `dashboard.html` from this skill's directory, replace `{{SKILL_NAME}}` in title

If workspace exists **with** `state.json`: read it and print pipeline status.

If workspace exists **without** `state.json` (created by a different tool): back up to `autoresearch-<skill>-prev/` and create a fresh workspace. The backup preserves prior artifacts as ground truth for Phase 2 comparison.

## Pipeline Status

Print at every session start:
```
AutoRefine: <name>
================================================================
Gulf 1: Comprehension
  Phase 1: Design Audit          [STATUS]
  Phase 2: Eval Audit             [STATUS]
  Phase 3: Error Analysis         [STATUS]  [N/M traces]
  >>> Gate: Approve taxonomy      [STATUS] <<<
Gulf 2: Specification  (v1.1)
  Phase 4-6: Judges               [NOT AVAILABLE YET]
Gulf 3: Generalization
  Phase 7: AutoResearch Loop      [STATUS]  [best score]
================================================================
```

---

## Phase 1: Design Audit

Assess structural quality against v2.0 skill design patterns.

**Read the target SKILL.md and score 4 dimensions:**

| Dimension | At Standard | Partial | Missing |
|-----------|-------------|---------|---------|
| **Gotchas** | 5-9 numbered failure modes, each names the failure + why + consequence | 1-4 or buried in other sections | None |
| **Voice** | "Do X because Y" throughout | Mix of instructional and descriptive | Purely descriptive |
| **Progressive Disclosure** | References tagged `Read when: [condition]`, folder structure | Some separation | Single flat file |
| **Scripts** (if any) | `__all__`, type hints, `Use when:` docstrings | Partial | No composability |

For each Partial/Missing dimension, write: what's wrong (quote actual text), recommended fix (concrete action), priority (HIGH/MEDIUM/LOW).

**Output:** Write `design-audit.md` in workspace with scores table + findings.
**State:** Mark phase 1 complete, advance to phase 2. For detailed rubric, read `references.md`.

---

## Phase 2: Eval Audit

Assess existing eval infrastructure, or document its absence.

**Check the target skill directory for:** eval-suite.md, evals.json, test fixtures, results files, autoresearch artifacts.

**If evals exist, audit against 6 categories:**
1. **Error analysis grounding** — Were evals built from observed failures, or brainstormed? Look for error-analysis-traces.md or failure taxonomy.
2. **Evaluator design** — Binary pass/fail? Or vague Likert scales? Are holistic evals bundling multiple failure modes?
3. **Judge validation** — Any TPR/TNR measurements? Golden dataset? Or untested judges?
4. **Train/test split** — Same fixtures for iteration AND measurement? (= data leakage)
5. **Labeled data** — How many labeled examples? Target: >50. Under 25 is critical gap.
6. **Maintenance** — Process for re-auditing after skill changes? Or "set and forget"?

**If no evals exist:** Document: "No eval infrastructure. Phase 3 builds the foundation."

### Enhancement: eval-audit (Hamel)
If `hamel_available` is true in state.json, invoke the `eval-audit` skill for deeper analysis. It runs the same 6 diagnostics above but with richer heuristics: flags class imbalance in metrics, detects stale analyses, and recommends specific next skills.

**How to invoke:** Provide the eval artifacts (eval-suite.md, results files, fixture paths) as context. The skill produces a structured findings report with problem title, status, explanation, and recommended fix for each diagnostic.

**Output:** Write `eval-audit-report.md` with findings and remediation priorities.
**State:** Mark phase 2 complete, advance to phase 3.

---

## Phase 3: Error Analysis

Close the Gulf of Comprehension. This is the most important phase. CANNOT BE AUTOMATED.

### Step 1: Prepare fixtures
Check existing test inputs in the skill directory. If fewer than 15 diverse inputs exist, help generate additional ones covering: different input types, lengths, quality levels, edge cases, at least 2 with planted flaws.

### Step 2: Run the skill
Invoke the target skill on each fixture (15-25 inputs). Save each output to `traces/trace-T01.md` through `traces/trace-T25.md`.

**Session-spanning skills:** Some skills run *during* an entire session (e.g., tracing, monitoring, coaching) rather than producing output on a single input. For these:
- Don't try to invoke the skill 15+ times — it's not a one-shot tool.
- Instead, create **synthetic output fixtures** that simulate what the skill would produce across different scenarios (clean session, error-heavy, short session, edge cases).
- Also include any **real outputs** from prior runs if they exist (e.g., from a previous autoresearch).
- The goal is the same: give the human diverse outputs to review. The generation method adapts to the skill type.

### Step 3: Human reviews outputs
Present traces in batches of 5. User can stop after any batch.

For each trace, ask:
- **Pass or Fail?** (overall quality judgment)
- **If Fail:** What went wrong? (free text — no pre-defined categories)
- **If Pass:** Anything surprising or borderline?

Record in `error-analysis-traces.md`:
```
| # | Fixture | Pass/Fail | Notes |
|---|---------|-----------|-------|
| T01 | fixture-name.md | FAIL | Location references too vague, used "throughout" |
```

**Mid-phase resume:** Track `traces_reviewed` in state.json. Next session says: "You reviewed N of M traces. Continue with T-XX?"

### Step 4: Build failure taxonomy
After all traces reviewed, cluster failure notes into categories. Let categories EMERGE — do NOT use existing eval categories as starting taxonomy. Present to user for approval.

Write `failure-taxonomy.md`:
```
## Failure Taxonomy: <skill>
1. [Category Name] — [description] — observed in N/M traces
2. ...
```

### Step 5: Generate initial eval suite
Convert top failure categories into binary evals in `eval-suite.md`:
```
EVAL 1: [Name]
Question: [specific yes/no question]
Pass: [what success looks like]
Fail: [what failure looks like]
Why: [which observed failure mode this catches]
```

### Gate: Gulf 1 Exit
Generate `gate-report-gulf-1.md`:
- Traces reviewed, fail rate, top failure categories
- Proposed eval suite (from Step 5)
- Decision: "Approve this taxonomy and eval suite to proceed?"

**STOP. Wait for user approval before Phase 4.**

### Enhancement: error-analysis (Hamel)
If `hamel_available` is true, invoke the `error-analysis` skill to structure the review process. Adaptations for skills vs. LLM pipelines:
- Hamel's skill expects ~100 production traces → use 20-25 fixture outputs instead
- Hamel's skill uses LLM-assisted clustering after ~30 traces → use manual clustering (our trace count is smaller)
- Hamel's skill recommends random/stratified/outlier sampling → use fixture diversity instead (we control inputs)
- The core protocol is the same: read every output → judge Pass/Fail → capture root cause (not explanation) → cluster into 5-10 categories → compute failure rates

### Enhancement: generate-synthetic-data (Hamel)
If `hamel_available` is true, use the `generate-synthetic-data` skill to prepare diverse fixtures in Step 1. It generates inputs via dimension-based tuples:
1. Define 3 failure-prone dimensions (e.g., Document Type × Quality Level × Domain)
2. Draft 20 tuples with user feedback
3. LLM generates more tuples, user validates
4. Convert tuples to natural language test inputs
This produces more systematically diverse fixtures than ad hoc generation.

**State:** Mark phase 3 complete. Record traces_reviewed and taxonomy summary.

---

## Phases 4-6: Specification (v1.1)

These phases close the Gulf of Specification. They will be added in v1.1 after validating Gulf 1 on a real skill. Summary of what they do:

- **Phase 4: Expand Inputs** — Use `generate-synthetic-data` (Hamel) or manual generation to build 30+ diverse fixtures with dimension-based tuples. Split into train/dev/test sets.
- **Phase 5: Write Judges** — Use `write-judge-prompt` (Hamel) to create binary LLM-as-judge evaluators grounded in the failure taxonomy from Phase 3. Each judge needs: task/criterion definition, Pass/Fail definitions, 3+ few-shot examples from train split, structured JSON output. Do NOT use dev/test examples as few-shots (data leakage).
- **Phase 6: Validate Judges** — Use `validate-evaluator` (Hamel) to calibrate each judge against a golden dataset. Run judge on dev split → measure TPR/TNR → inspect disagreements → refine prompt → repeat until TPR >90% AND TNR >90%. Final measurement on held-out test split. Apply Rogan-Gladen correction for production estimates.

**Gulf 2 Gate:** After Phase 6, user must approve judges before autoresearch begins.

---

## Phase 7: AutoResearch Loop

The Karpathy-style mutation-test-keep/discard cycle. Requires a populated `eval-suite.md` (from Phase 3 or manually created).

### The Loop

```
1. Read eval-suite.md for criteria and fixtures
2. Run baseline: invoke skill on all fixtures, score against all evals → Experiment 0
3. Record in results.json, results.tsv, changelog.md
4. LOOP:
   a. Analyze failures: which evals fail on which fixtures? What pattern?
   b. Hypothesize: what ONE change to SKILL.md would fix the failing pattern?
   c. Mutate: make the change to a copy (<skill>-optimized.md)
   d. Test: run mutated skill on all fixtures, score against all evals
   e. Decide:
      - Score improved → KEEP. Update baseline. Record in changelog.
      - Score same/worse → DISCARD. Revert. Record in changelog.
   f. Update results.json, results.tsv, dashboard refreshes automatically
   g. Repeat until: all evals pass, or budget exhausted (default: 5 experiments)
```

### Changelog Format
```markdown
## Experiment N — [keep/discard]
**Score:** X/Y (Z%)
**Change:** [what was mutated in the skill]
**Reasoning:** [why this change should help]
**Result:** [what happened — which evals flipped]
**Failing outputs:** [remaining failures, or "None"]
```

### Results Schema
Append to `results.json.experiments[]`:
```json
{"id":N,"score":X,"max_score":Y,"pass_rate":Z,"status":"keep|discard|baseline","description":"...","changes":[{"type":"added|modified|removed","location":"section or line ref","snippet":"the actual text changed"}]}
```
The `changes` array captures what was mutated — the dashboard renders these as expandable details per experiment. Keep snippets short (1-3 lines each). The `location` field says WHERE in the skill, `snippet` shows WHAT changed.
Append to `results.tsv`:
```
N\tX\tY\tZ%\tkeep\tdescription
```

### Key Rules
- **One mutation per experiment.** Multi-variable changes make attribution impossible.
- **Score the MUTATED copy, not the original.** Keep original SKILL.md untouched as the baseline.
- **If score improves, the mutated copy becomes the new baseline** for next experiment.
- **Budget cap: ask the user before starting.** Present three options:
  - **Quick (3 experiments)** — for low-priority skills or token-conscious users. Targets the 2-3 worst evals only.
  - **Standard (5 experiments)** — default. Covers the major failure categories.
  - **Deep (8-10 experiments)** — for critical skills the user wants to fully optimize. Extends into edge cases and sparse-trace handling.

  Say: "How many experiments do you want to run? Quick (3), Standard (5), or Deep (8-10)? Quick is best for low-priority skills, Deep for skills you use daily."
- **Target baseline 60-80%.** If baseline >90%, evals are too easy — harden them first.

### Enhancement: skill-creator subagents
If skill-creator is available, use its specialized subagents to strengthen the loop:

**Grader** — After each experiment run, dispatch the grader subagent with the eval expectations and skill output. It returns structured pass/fail verdicts with evidence, verifies claims from the output, AND critiques the evals themselves (flags assertions that would pass bad outputs). Provide: expectations list, output files, transcript.

**Comparator** — For rigorous A/B testing between the baseline and mutated skill, dispatch the comparator with both outputs (blinded — it doesn't know which is which). It scores on content + structure rubrics and picks a winner. Use when score deltas are small and you need confidence the mutation actually helped.

**Analyzer** — After the comparator picks a winner, dispatch the analyzer with both skills + transcripts. It explains WHY the winner won and produces prioritized improvement suggestions. Use to inform your next mutation hypothesis.

### Serving the Dashboard
```bash
cd autoresearch-<skill>/
python3 -m http.server 8080
# Open http://localhost:8080/dashboard.html
```

**State:** Update after each experiment. Mark phase 7 complete when loop ends.

---

## Gotchas

1. **Don't skip Gulf 1.** A 100% score on narrow evals is an artifact, not evidence of quality. Comprehension first, always.

2. **Error analysis cannot be automated.** Phase 3 requires the human to read outputs. An LLM doing it for you is comprehension theater.

3. **100% on narrow evals ≠ quality.** Target 60-80% baseline. If baseline passes everything, harden your evals.

4. **Let categories emerge.** In Phase 3, don't start with existing eval categories. Fresh eyes, fresh taxonomy.

5. **One mutation per experiment.** Multi-variable changes make attribution impossible.

6. **Keep rate matters more than final score.** 60% → 85% through 4 keeps out of 10 teaches more than 95% → 100% in 1.

7. **High Phase 3 fail rates are expected.** A 60-100% fail rate in error analysis means your fixtures are diverse and your reviewer is rigorous — this is healthy. It produces a rich failure taxonomy that drives meaningful evals. A low fail rate (<30%) usually means the fixtures are too easy or the reviewer is too generous.

8. **Dashboard needs HTTP serving.** `python3 -m http.server 8080` from the autoresearch directory. Direct `file://` won't work (CORS). Dashboard also requires internet for Chart.js CDN — if blocked, download `chart.umd.min.js` locally and update the script src.

9. **Never run two sessions on the same skill.** state.json has no locking. Parallel sessions will overwrite each other's results, producing corrupted state and lost experiments.

10. **"Invoke" means "read and follow."** When this skill says "invoke eval-audit" or "dispatch the grader subagent," it means: read that skill's SKILL.md and follow its instructions inline. Not all agents support direct skill invocation — reading the instructions works everywhere.

## When AutoRefine Doesn't Help

These are failure modes of autorefine itself. Know when to step back.

1. **Everything scores "At Standard" but the skill is still bad.** The design audit checks structural patterns (Gotchas, voice, disclosure). A well-structured skill can still produce wrong outputs. If Phase 1 passes but users complain, skip to Phase 3 (error analysis) — the problem is in the logic, not the structure.

2. **Phase 3 fail rate is very low (<20%).** This usually means your fixtures aren't diverse enough, or you're being too generous in review. Add harder fixtures — edge cases, adversarial inputs, inputs from different domains. If fail rate is still low, the skill may actually be good and autorefine isn't needed.

3. **AutoResearch loop plateaus — no improvement after 3+ experiments.** Your evals may not be discriminating enough (too easy or too binary to capture nuance). Or the skill's failure mode requires architectural change, not prompt mutation. Consider: rewriting the skill from scratch vs. mutating it, or adding new eval dimensions you haven't measured.

4. **Fixtures don't represent real usage.** If all your test inputs are clean, well-formatted examples but real users send messy, ambiguous inputs, the autoresearch loop optimizes for the wrong surface. Include at least 3-5 "ugly" real-world inputs in your fixture set.

## References

- `references.md` — Read when: user asks "why this order", "why can't I skip to autoresearch", or needs the v2.0 audit rubric details
登录后下载/点赞/收藏 ❤ 3 | ★ 0
autorefine

目录结构

SKILL.md

举报内容

提示
