Improve a skill through iterative executor runs, quantitative scoring, targeted fixes, and reruns until stop criteria are met. Inspired by Karpathy's autoresearch.

Skill: Skill Autoresearch

When to Use

After building or significantly updating a skill, to validate it works end-to-end
When a skill has known confusion points or failure modes
Before promoting a skill from draft to production use
When onboarding a skill written by someone else (stress-test before trusting)
Periodically, to catch drift between SKILL.md instructions and actual tool behavior

How It Works

                         ┌──────────────────────────┐
                         │     SETUP (one-time)     │
                         │                          │
                         │  1. Choose target skill  │
                         │  2. Create test brief    │
                         │  3. Define stop criteria  │
                         └──────────┬───────────────┘
                                    │
              ┌─────────────────────▼─────────────────────┐
              │                                            │
              │   ITERATION LOOP                           │
              │                                            │
              │   ┌──────────────────────────────────┐     │
              │   │ Executor Agent (worktree)         │     │
              │   │ - Reads target SKILL.md           │     │
              │   │ - Follows procedure step by step  │     │
              │   │ - Uses test brief as input         │     │
              │   │ - Reports confusion log per step  │     │
              │   └──────────────┬─────────────────────┘    │
              │                  │                          │
              │   ┌──────────────▼─────────────────────┐    │
              │   │ Score (orchestrator)                │    │
              │   │ - Average clarity across steps      │    │
              │   │ - Tool/script success (0/1)         │    │
              │   │ - Output quality metrics            │    │
              │   │ - Compare to stop criteria          │    │
              │   └──────────────┬─────────────────────┘    │
              │                  │                          │
              │         ┌────────▼────────┐                 │
              │         │ Criteria met?    │                 │
              │         └───┬─────────┬───┘                 │
              │          No │         │ Yes                  │
              │   ┌────────▼──────┐  │                      │
              │   │ Fix skill     │  │                      │
              │   │ files based   │  │                      │
              │   │ on findings   │  │                      │
              │   └────────┬──────┘  │                      │
              │            │         │                      │
              └────────────┘    ┌────▼─────────────────┐    │
                                │ DONE                  │    │
                                │ Report final scores   │    │
                                │ + changelog           │    │
                                └───────────────────────┘

Core Concepts

The Executor Agent Is the Test

The key insight: instead of writing test scripts, you launch an agent that follows the skill procedure as a real user would. The agent's confusion IS the test signal. If it gets confused, the skill's instructions are unclear. If the output is wrong, the procedure has a gap.

This is mandatory, not optional. Do not replace the executor-agent run with a local checklist, fixture smoke test, or manual paraphrase of what the executor would have done.

Confusion Score Is the Primary Metric

Each step in the target skill gets a clarity rating (1-10). The average across all steps is the primary optimization target. This is the equivalent of Karpathy's val_bpb — a single quantitative number that improves monotonically as the skill gets clearer.

Worktree Isolation Keeps Experiments Clean

Each executor runs in a git worktree, so it can create files, run scripts, and make changes without affecting the working directory. Failed experiments are discarded automatically. This is analogous to Karpathy's branch-per-experiment pattern.

This is mandatory for every executor run. Do not run the executor in the main worktree.

3-5 Iterations Is Typical

Skill procedures converge fast — unlike ML hyperparameter search (100+ experiments), prose clarity improvements plateau after 3-5 rounds. If you're still finding major issues after 5 iterations, the skill may need architectural changes, not instruction tweaks.

Procedure

Step 1: Choose the target skill and read it

Identify the skill to test. Read:

The target skill's SKILL.md
Any templates, scripts, or reference files it uses
Any existing output examples (to understand what "good" looks like)

Note the skill's:

Steps: How many steps does the procedure have?
Modes: Does it have multiple modes (create/update/import)?
Tools/scripts: Does it call external scripts? Which ones?
Outputs: What files does it produce?

Step 2: Ask about persistence before run

Before creating briefs or fixtures, ask the user whether they want persistence.

Use this decision flow:

Ask: Do you want to persist test files?
If yes, ask: In which folder? Default path is workdir/autoresearch-evals/{skill-name}/
If no, keep everything runtime-only and save nothing permanently.

Do not assume persistence without asking first.

Step 3: Create a test brief

A test brief is a simulated set of "source materials" that an executor agent uses as input. It replaces the real user.

Requirements:

Must contain enough information to fill at least 80% of the skill's expected output
Should include some intentional gaps (to test how the skill handles missing info)
Should include at least one edge case (unusual value, ambiguous input, boundary condition)
Must be self-contained — the executor should not need external resources

Default storage: do not save the brief permanently. Keep it in runtime only, such as runtime/.tmp/skill-autoresearch/{skill-name}/{brief-name}.md, or inline it in the executor prompt if small.

If persistence is explicitly requested by the user after the preflight question:

Save briefs and fixtures in the folder the user asked for.
Also save result.json in that same folder.
If the user wants persistence but gives no folder, use workdir/autoresearch-evals/{skill-name}/.
Use shorten Scenario Name for filenames. Good: update-existing-file.md, ambiguous-input.md. Bad: 260417-brief.md, tmp.md.

Important split:

By default, briefs, fixtures, scores, and run output are ephemeral. Keep them in runtime or the isolated worktree only.
Persist briefs or fixtures only when the user explicitly asks to keep them.
If fixtures are persisted, store them next to the persisted brief in a user-chosen folder, or under workdir/autoresearch-evals/{skill-name}/fixtures/ / scripts/ if no folder was specified.
Executor run output is always ephemeral. Materialize fixtures into the isolated worktree for that run; do not reuse prior run output directories.
Do not spend tokens re-describing large file trees if a fixture or setup script can express the same scenario once.

Score tracking:

Do not store last score inside the brief itself. The brief is input; scores are run output.
By default, keep scores in the run report only; do not write a persistent score file.
If persistence was requested, store scores separately in result.json next to the persisted brief and fixtures.
Record at minimum: brief name, latest score, best score, date, target skill version tested, notes.

Template:

# Test Brief: [Scenario Name]

> Simulated input for testing the [skill-name] skill.
> The executor agent treats this as what a real user would provide.

## Scenario
[Describe the test scenario — what is the user trying to accomplish?]

## Source Materials
[The actual content the executor will use as input — colors, text, data, files, etc.]

## Simulated User Behavior
[How the executor should respond when the skill asks for user input]
- When asked to confirm: [confirm / reject / partial]
- When asked to fill gaps: [provide values / skip / done]
- When asked to choose: [which option]

## Expected Output
[What the output should approximately contain — used by reviewer for validation]

## Edge Cases Included
[List the intentional edge cases and what correct handling looks like]

If the scenario needs non-trivial starting files and persistence was requested, prefer this pattern inside the brief:

## Fixture Source

- Materialize from `./fixtures/[fixture-name]/`
- Or run `./scripts/setup-[scenario].sh` in the isolated worktree

If persistence was not requested, describe the setup inline or create it directly inside the isolated worktree for that run.

Step 4: Define stop criteria

Every autoresearch loop needs explicit stop criteria. Without them, you'll iterate forever on diminishing returns.

Default stop criteria (use these unless the skill warrants different thresholds):

Metric	Target	How Measured
Instruction clarity	>= 8.5/10 average	Executor confusion log
Script/tool success	100% (all pass)	Exit codes from skill's scripts
Output completeness	>= 75% of expected fields	Parser/validator output or manual count
Mode coverage	All modes tested	At least one test per mode
No blocking issues	0 remaining	Executor reports no confusion >= 3 steps below 7/10
Brief leakage	0 brief-specific literals copied into the skill	Reviewer diff inspection

Custom metrics for skills with measurable output:

Coverage percentage (for skills that generate structured data)
Validation pass rate (for skills that produce parseable output)
Accessibility scores (for design/brand skills)
Test pass rate (for skills that produce code)

Step 5: Run the iteration loop

For each iteration:

5a. Launch the executor agent

This step is mandatory for every brief you count as tested.

Use the Agent tool with:

isolation: "worktree" — so changes don't affect the main workspace
mode: "auto" — so the executor can create files and run scripts

Requirements:

launch a separate executor subagent per brief or per mode under test; do not collapse multiple briefs into one executor report
each executor must perform end-to-end execution of the target skill as written, not only fixture setup or syntax checks
each executor must return a confusion log covering every step in the target skill's procedure
do not mark a brief as tested if the executor skipped the confusion log or did not complete the full procedure

Before launching:

If persisted fixtures or scripts exist, copy them into the worktree or run them there.
Otherwise create the scenario setup directly in the isolated worktree for that run only.
Start from the brief and fixture inputs only; do not preload prior execution artifacts.
Keep skill fixes generic. The goal is to improve the skill, not to optimize for one scenario's wording.

Executor prompt template (copy and customize):

You are a test executor for iteration N of an autoresearch loop. Your job is to follow
the [SKILL NAME] skill procedure EXACTLY as written. Read all files fresh — do not
assume you know the contents.

## Your Task

1. Read the SKILL.md at [path to target SKILL.md]
2. Read any templates/reference files the skill uses
3. Read the test brief at [path to test brief]
4. Materialize any referenced fixture or run any referenced setup script inside the isolated worktree
5. Follow the procedure step by step in [MODE] mode
6. Treat the brief as test input, not as source text to transplant into the skill
7. Complete ALL steps including validation/output generation

## Simulated User Behavior

[Copy from test brief]

## CRITICAL: Confusion Report

Rate each step on clarity (1-10). For EACH step, note:
- Was the instruction clear? (yes/no)
- Did you have to guess or interpret anything ambiguous?
- Were there contradictions between instructions and templates?
- Did any guidance added since the last iteration help?
- Did any proposed fix look specific to this brief rather than general to the skill?

## Output Format

### 1. Execution Result
- Skill completed end-to-end? (yes/no)
- Tool/script outputs (full text of validation, coverage, errors)
- Files created

### 2. Confusion Log
Step 1: Clarity X/10
- Issue: ...
(Continue for all steps)

### 3. Issues Found
- Template issues
- Script/tool issues
- Procedure gaps
- Brief-specific leakage risk

### 4. Overall Assessment
- What worked well
- What is still confusing
- Suggestions for improvement

5b. Score the results

After the executor returns, compute scores against stop criteria:

Metric	Score	Pass?
Instruction clarity	[avg of step scores]	[>= target?]
Script/tool success	[pass/fail]	[all pass?]
Output completeness	[%]	[>= target?]
Mode coverage	[modes tested / total modes]	[all covered?]
Blocking issues	[count of steps < 7/10]	[0?]
Brief leakage	[count]	[0?]

5c. Fix or stop

If all criteria met: Stop the loop. Report final scores and changelog.

If criteria not met: Analyze the executor's confusion log and fix the target skill files:

Low-clarity steps → rewrite instructions for that step
Script failures → fix the script
Template confusion → clarify field descriptions, add notes
Procedure gaps → add missing steps or guidance
Contradictions → resolve in favor of the more specific instruction
Brief-specific wording or naming leakage → replace with generalized language grounded in the skill's own domain and source files, not the test brief's incidental labels

Bump the skill version after each round of fixes (patch for minor clarifications, minor for new steps/guidance).

Then launch the next iteration.

Step 6: Test additional modes

If the target skill has multiple modes (create/update, import/export, etc.), test each mode at least once. You can test secondary modes in the same iteration as the final create-mode pass.

For update mode specifically:

Start from the output of a successful create-mode run
Apply a targeted change (swap a value, add a field, remove a section)
Verify the change propagated correctly through any downstream tools

Step 7: Report final results

When the loop converges, present:

Convergence summary:

Autoresearch: [skill name]
Iterations: N
Clarity: [start] → [end]
Changes: [count of edits to SKILL.md, template, scripts]
Modes tested: [list]
Stop criteria: ALL MET

Changelog (what was fixed per iteration):

Iteration 1 (clarity X.X):
- [finding] → [fix applied]
- [finding] → [fix applied]

Iteration 2 (clarity X.X):
- [finding] → [fix applied]

When reporting a persisted run, also include enough executor evidence to audit the claim:

which executor subagent handled each brief
whether it ran in an isolated worktree
whether it produced a per-step confusion log

Version bumps:

SKILL.md: v[old] → v[new]
Template: v[old] → v[new] (if applicable)

Guidance

When to run more iterations

Any step below 7/10 → must fix and re-test
Average below 8.5/10 → should fix and re-test
Script/tool failures → must fix and re-test
Mode not tested → must test
Any brief-specific leakage into the skill → must fix and re-test

When to stop

Average >= 8.5/10 AND no step below 7/10 AND all scripts pass AND all modes tested AND no brief leakage
OR: 5 iterations completed with diminishing returns (< 0.3 improvement per round)
OR: Remaining issues are inherent to the skill's domain (e.g., a brand's accent color will always fail AA contrast — that's not the skill's fault)

Common patterns from experience

Token/field mapping tables are the highest-ROI addition. If a skill expects structured input, provide a mapping guide from common source terminology to the skill's vocabulary.
System-documentation columns (like fallback/default values in tables) must be explicitly marked as read-only. Agents will try to edit them.
Exit-code behavior of called scripts should be documented in the procedure. A non-zero exit that's informational (not fatal) will confuse agents into thinking the step failed.
Metadata/coverage fields that depend on tool output need a write-back step at the end of the procedure. Don't say "filled automatically" when the agent has to do it.
Sequential vs. parallel tool execution should be called out when one tool's exit code can cancel the other.
Brief contamination is a real failure mode. If a test brief uses example names, labels, folder names, or edge-case phrasing, do not copy those literals into the skill unless they are truly part of the skill's permanent vocabulary. Convert findings into general rules.
Default to disposable runs. Persist fixtures only if the user asks. Reusing previous run output hides bugs.

Notes

This skill is orchestrator-level work — the agent running this skill is the researcher, not the executor. The executor is a subagent.
Default behavior: save nothing permanently. Keep briefs, fixtures, and scores in runtime only.
If the user explicitly asks to keep eval assets, save them in the folder the user requested.
If the user asks to keep eval assets but does not name a folder, use autoresearch-evals/{skill-name}/ and keep {brief-name}.md, fixtures/, scripts/, and result.json inside it.
Scratch files, executor output, and temporary materialized fixture copies belong in the isolated worktree or runtime/.tmp/.
Keep regression history separate from briefs. If persistence was requested, write it to result.json.
Worktree branches from executor agents are auto-cleaned if no changes were made. If changes persist, the worktree path is returned in the result.
The skill produces no deliverable — its output is the improved target skill itself.