Skill Creator

Create new skills and iteratively improve them. The lifecycle:

1. Decide what the skill should do and how
2. Draft the skill
3. Run test prompts through claude-with-the-skill
4. Evaluate results qualitatively and quantitatively with the user
5. Rewrite based on feedback
6. Repeat until satisfied, then expand the test set
7. Package and deliver

Determine where the user is in this lifecycle and help them progress. Starting from scratch ("I want a skill for X") -- help narrow scope, draft, test, iterate. Already have a draft -- go straight to eval/iterate. Just want a better description -- jump to Description Optimization.

Be flexible: if the user says "just vibe with me, no evals," do that.

Available Resources

Before starting, verify what is available in the environment. The full workflow requires these bundled resources:

agents/grader.md

agents/comparator.md

agents/analyzer.md

references/schemas.md

references/eval-patterns.md

references/troubleshooting.md

assets/eval_review.html

scripts/aggregate_benchmark

python -m scripts.aggregate_benchmark

scripts/run_loop

python -m scripts.run_loop

eval-viewer/generate_review.py

If resources are missing: The core loop (draft -> test -> review -> improve) still works without them -- grade inline and present results in-conversation instead of the browser viewer. Note what is unavailable to the user and adapt.

Communicating with the user

The skill creator serves people across a wide range of technical backgrounds. Non-developers are opening terminals for the first time because Claude makes things possible; experienced engineers are also common.

Pay attention to context cues:

• "evaluation" and "benchmark" are borderline but generally OK
• Avoid "JSON" or "assertion" without a brief explanation unless the user has already used those terms

Err toward clarity. A short inline definition never hurts.

Creating a skill

Capture Intent

The current conversation may already contain the workflow to capture. If the user says "turn this into a skill" or "make this repeatable," mine the conversation history first:

• What tools were used, in what sequence?
• What corrections did the user make?
• What were the input and output formats?
• What would a different user need to know to reproduce this?

Fill gaps with the user and confirm before drafting.

If starting from scratch, establish:

1. What should this skill enable Claude to do?
2. When should this skill trigger? (what user phrases, contexts, task types)
3. What is the expected output -- files, structured data, prose, or a workflow?
4. Should test cases be set up? Skills with verifiable outputs (file transforms, code generation, fixed workflows) benefit from them. Skills with subjective outputs (writing style, creative direction) often do not. Suggest the appropriate default based on skill type, but let the user decide.

Interview and Research

Ask about edge cases, input/output examples, success criteria, and dependencies before writing test prompts. If useful MCPs are available (docs search, similar skill lookup), research in parallel via subagents. Come prepared so the interview is lightweight.

Write the SKILL.md

Based on the interview, compose the skill. Every skill has:

• name: Identifier (kebab-case)
• description: The primary triggering mechanism. Describe what the skill does AND when to use it. All "when to trigger" information goes here, not in the body. Apply the "pushy principle" -- err toward including trigger contexts rather than omitting them. Example: instead of "Builds REST APIs," write "Builds REST APIs. Invoke whenever users ask about endpoints, routes, HTTP methods, API design, or want to add a backend -- even if they don't say 'REST.'"
• compatibility: Required tools or dependencies (optional, rarely needed)
• skill body: Instructions, workflow, examples, references

Skill output template

Use this as the starting structure -- fill in or remove sections as needed:

---
name: skill-name
description: >
  [What this skill does and when to trigger it. One to three sentences.
  Apply the pushy principle: include adjacent trigger contexts.]
---

# [Skill Name]

[One paragraph: what this skill enables and when to use it.]

## [Core workflow or main section]

[Instructions in imperative form. Explain the *why* behind each step.]

## Output format

[Concrete template for what this skill produces.]

## Examples

**Example 1:**
Input: [realistic user prompt]
Output: [expected result]

## Reference files

- Read `references/advanced.md` when [specific situation]
- Run `scripts/transform.py` for [repeatable operation]

Anatomy of a skill directory

skill-name/
+-- SKILL.md (required)
+-- Optional bundled resources:
    +-- scripts/    - Scripts for deterministic/repetitive tasks
    +-- references/ - Docs loaded into context as needed
    +-- assets/     - Templates, icons, fonts

Progressive loading:

1. Metadata (name + description) -- always in context
2. SKILL.md body -- in context when the skill triggers (<500 lines ideal)
3. Bundled resources -- loaded as needed, scripts can run without loading

Keep SKILL.md under 500 lines. If approaching that limit, add a layer of hierarchy with clear pointers to reference files. For reference files over 300 lines, add a table of contents.

Domain organization -- when a skill covers multiple frameworks/variants:

cloud-deploy/
+-- SKILL.md (workflow + selection logic)
+-- references/
    +-- aws.md
    +-- gcp.md
    +-- azure.md

Claude reads only the relevant reference file.

Writing patterns

Output templates -- be concrete:

## Report structure
ALWAYS use this template:
# [Title]
## Executive summary
## Key findings
## Recommendations

Examples -- include realistic ones:

## Commit message format
**Example:**
Input: Added user authentication with JWT tokens
Output: feat(auth): implement JWT-based authentication

Writing style -- use imperative form. Explain the why rather than issuing rules. Avoid ALWAYS/NEVER in all caps where the reasoning can be explained instead -- models respond better to understanding than mandates. Write a draft, then read it fresh and improve it.

Security -- skills must not contain malware, exploit code, or anything that would surprise the user if they read the description. Do not create skills designed to facilitate unauthorized access or data exfiltration. Roleplay skills are fine.

Test Cases

After drafting, write 2-3 realistic test prompts -- things a real user would actually type. Share them: "Here are a few test cases I'd like to try. Do these look right, or do you want to add more?"

Save to

evals/evals.json

{
  "skill_name": "example-skill",
  "evals": [
    {
      "id": 1,
      "prompt": "User's task prompt",
      "expected_output": "Description of expected result",
      "files": []
    }
  ]
}

See

references/schemas.md

assertions

Running and evaluating test cases

This is one continuous sequence -- do not stop partway through. Do NOT use

/skill-test

Put results in

<skill-name>-workspace/

Workspace layout

<skill-name>-workspace/
+-- iteration-1/
|   +-- eval-<name>/                    <-- one per test case (descriptive name)
|   |   +-- eval_metadata.json          <-- eval ID, prompt, assertions
|   |   +-- with_skill/
|   |   |   +-- run-1/
|   |   |       +-- outputs/            <-- files the subagent produced
|   |   |       |   +-- metrics.json
|   |   |       +-- timing.json         <-- from task notification
|   |   |       +-- grading.json        <-- written by grader agent
|   |   +-- without_skill/              <-- or old_skill/ when improving
|   |       +-- run-1/
|   |           +-- outputs/
|   |           +-- timing.json
|   |           +-- grading.json
|   +-- benchmark.json                  <-- written by aggregate_benchmark.py
|   +-- benchmark.md
+-- iteration-2/
    +-- ...

Step 1: Spawn all runs (with-skill AND baseline) in the same turn

For each test case, spawn two subagents simultaneously -- one with the skill, one without. Do not do with-skill runs first and baseline runs later. Launch everything at once so results arrive around the same time.

With-skill run:

Execute this task:
- Skill path: <path-to-skill>
- Task: <eval prompt>
- Input files: <eval files if any, or "none">
- Save outputs to: <workspace>/iteration-<N>/eval-<name>/with_skill/run-1/outputs/
- Outputs to save: <what the user cares about>

Baseline run -- depends on context:

• New skill: no skill at all. Same prompt, save to

  <workspace>/iteration-<N>/eval-<name>/without_skill/run-1/outputs/

• Improving existing skill: snapshot the old version first (

  cp -r <skill-path> <workspace>/skill-snapshot/

  ), point baseline at snapshot, save to

  old_skill/run-1/outputs/

Write

eval_metadata.json

<workspace>/iteration-N/eval-<name>/eval_metadata.json

{
  "eval_id": 0,
  "eval_name": "descriptive-name-here",
  "prompt": "The user's task prompt",
  "assertions": []
}

Step 2: While runs are in progress, draft assertions

Do not just wait for runs to finish -- use this time productively. Draft quantitative assertions for each test case and explain them to the user. If assertions already exist in

evals/evals.json

Good assertions are objectively verifiable and have descriptive names -- someone glancing at benchmark results should immediately understand what each one checks. For subjective skills (writing style, design quality), skip assertions and rely on qualitative review. See

references/eval-patterns.md

Update

eval_metadata.json

evals/evals.json

Step 3: As runs complete, capture timing data

When each subagent task completes, save timing data immediately to

timing.json

with_skill/run-1/timing.json

{
  "total_tokens": 84852,
  "duration_ms": 23332,
  "total_duration_seconds": 23.3
}

This data only comes through task completion notifications -- capture it then.

Step 4: Grade, aggregate, and launch the viewer

GENERATE THE EVAL VIEWER BEFORE EVALUATING INPUTS YOURSELF. Get outputs in front of the human first.

Once all runs complete:

1. Grade each run -- spawn a grader subagent for each run directory (both with_skill and without_skill/old_skill get their own grader). Use this prompt:

   You are a grader agent. Read agents/grader.md from <skill-creator-path>
   to load your full instructions, then grade this eval run:
   
   - Expectations: ["assertion 1", "assertion 2", ...]
   - Transcript path: <workspace>/iteration-N/eval-<name>/with_skill/run-1/transcript.md
   - Outputs dir:    <workspace>/iteration-N/eval-<name>/with_skill/run-1/outputs/
   
   Save grading.json to <workspace>/iteration-N/eval-<name>/with_skill/run-1/grading.json.
   

   Required field names in

   grading.json

   :

   text

   ,

   passed

   ,

   evidence

   (not

   name

   /

   met

   /

   details

   -- the viewer depends on these exact names). For assertions that can be checked programmatically, write and run a script.

2. Aggregate -- run:

   python -m scripts.aggregate_benchmark <workspace>/iteration-N --skill-name <name>
   

   Produces

   benchmark.json

   and

   benchmark.md

   . List with_skill before baseline. If generating

   benchmark.json

   manually, see

   references/schemas.md

   for the exact schema the viewer expects.

3. Analyst pass -- read the benchmark data and surface patterns the aggregate stats might hide: non-discriminating assertions (always pass regardless of skill), high-variance evals (possibly flaky), time/token tradeoffs. See

   agents/analyzer.md

   ("Analyzing Benchmark Results" section).

4. Launch the viewer:

   nohup python <skill-creator-path>/eval-viewer/generate_review.py \
     <workspace>/iteration-N \
     --skill-name "my-skill" \
     --benchmark <workspace>/iteration-N/benchmark.json \
     > /dev/null 2>&1 &
   VIEWER_PID=$!
   

   For iteration 2+, add

   --previous-workspace <workspace>/iteration-<N-1>

   .

   No display / headless: use

   --static <output_path>

   to write a standalone HTML file. The "Submit All Reviews" button downloads

   feedback.json

   -- copy it into the workspace directory when done.

5. Tell the user: "I've opened the results in your browser. 'Outputs' tab lets you review each test case and leave feedback. 'Benchmark' tab shows the quantitative comparison. Come back when you're done."

What the user sees

Outputs tab: one test case at a time -- prompt, output, previous output (collapsed, iteration 2+), formal grades (collapsed), feedback textbox, previous feedback.

Benchmark tab: pass rates, timing, token usage per configuration, per-eval breakdowns, analyst observations.

Navigation: prev/next buttons or arrow keys. "Submit All Reviews" saves

feedback.json

Step 5: Read the feedback

{
  "reviews": [
    {"run_id": "eval-0-with_skill", "feedback": "chart is missing axis labels"},
    {"run_id": "eval-1-with_skill", "feedback": ""},
    {"run_id": "eval-2-with_skill", "feedback": "perfect, love this"}
  ],
  "status": "complete"
}

Empty feedback means the user was satisfied. Focus improvements on cases with specific complaints. Kill the viewer when done:

kill $VIEWER_PID 2>/dev/null

Improving the skill

This is the heart of the loop. Tests have been run, the user has reviewed -- now make the skill better.

How to think about improvements

Generalize from feedback, do not overfit to examples. The skill will run across countless different prompts. A skill that works only for the test examples is useless. Rather than fiddly constraints, try different metaphors, different working patterns, different framings -- cheap to try, potentially transformative.

Keep the prompt lean. Read the transcripts, not just final outputs. If the skill makes the model waste time on unproductive steps, remove those instructions. Deadweight is not neutral -- it adds noise and slows execution.

Explain the why, not just the what. Models have good theory of mind. They perform better with reasoning than rules. If ALWAYS or NEVER in all caps appears, try instead to explain why the thing matters. "Show the loading state before fetching so users aren't left wondering if anything happened" outperforms "ALWAYS show loading state."

Bundle repeated work. If all test case transcripts show the subagent writing the same helper script, write it once, put it in

scripts/

Take time here. Read the draft revision with fresh eyes. Try to genuinely understand what the user wants and what is blocking them, then transmit that understanding into the instructions.

The iteration loop

1. Apply improvements to the skill
2. Rerun all test cases into

   iteration-<N+1>/

   , including baselines
3. Launch the viewer with

   --previous-workspace

   pointing at the previous iteration
4. Wait for the user to review
5. Read feedback, improve again, repeat

Stop when: the user says they are happy, feedback is all empty, or meaningful progress has stopped.

Advanced: Blind comparison

For rigorous comparison between two skill versions, use the blind comparison system: read

agents/comparator.md

agents/analyzer.md

Description Optimization

The

description

Step 1: Generate trigger eval queries

Create 20 queries -- a mix of should-trigger and should-not-trigger. Save as JSON:

[
  {"query": "the user prompt", "should_trigger": true},
  {"query": "another prompt", "should_trigger": false}
]

Make queries realistic and specific. Include file paths, personal context, column names, company names, casual speech, typos, abbreviations, varying lengths. Focus on edge cases, not clear-cut cases.

Bad:

"Format this data"

"Create a chart"

Good:

"ok so my boss just sent me this xlsx file (its in my downloads, called something like 'Q4 sales final FINAL v2.xlsx') and she wants me to add a column that shows the profit margin as a percentage. The revenue is in column C and costs are in column D i think"

Should-trigger queries (8-10): Cover different phrasings of the same intent. Include cases where the user clearly needs the skill but does not name it. Include uncommon use cases and cases where this skill competes with another but should win.

Should-not-trigger queries (8-10): The most valuable are near-misses -- queries that share keywords but need something different. Adjacent domains, ambiguous phrasing where keyword-matching would fire but should not. Do not use obviously irrelevant negatives -- they test nothing.

Step 2: Review with user

1. Read

   assets/eval_review.html

2. Replace

   __EVAL_DATA_PLACEHOLDER__

   with the JSON array (no quotes -- JS variable assignment)
3. Replace

   __SKILL_NAME_PLACEHOLDER__

   and

   __SKILL_DESCRIPTION_PLACEHOLDER__

4. Write to

   /tmp/eval_review_<skill-name>.html

   and open it
5. User can edit queries, toggle should-trigger, add/remove entries, then "Export Eval Set"
6. Check Downloads for the most recent

   eval_set.json

Step 3: Run the optimization loop

python -m scripts.run_loop \
  --eval-set <path-to-trigger-eval.json> \
  --skill-path <path-to-skill> \
  --model <model-id-powering-this-session> \
  --max-iterations 5 \
  --verbose

Use the model ID from the system prompt so the test matches what the user actually experiences. Run in background and tail output periodically for updates.

The loop: 60/40 train/test split, evaluate current description (3 runs per query for variance), propose improvements from failures, re-evaluate, pick best by test score (not train, to avoid overfitting), repeat up to 5 iterations.

How triggering works: Claude sees skills as

(name, description)

Step 4: Apply the result

Take

best_description

Package and Present

If the

present_files

python -m scripts.package_skill <path/to/skill-folder>

Present the resulting

.skill

Claude.ai-specific instructions

Same core workflow, different mechanics:

Running test cases: No subagents. Read SKILL.md, follow instructions to complete each test prompt directly, one at a time. Skip baselines. Less rigorous (the skill author is also running it), but useful as a sanity check -- human review compensates.

Reviewing results: If no browser is available, present results directly in conversation. For file outputs (.docx, .xlsx), save to filesystem and tell the user where to find them. Ask for feedback inline.

Benchmarking: Skip quantitative benchmarking -- baseline comparisons are not meaningful without subagents. Focus on qualitative user feedback.

Description optimization: Requires

claude -p

Blind comparison: Requires subagents. Skip.

Updating an existing skill:

• Preserve the original name (directory name and

  name

  frontmatter field unchanged)
• The installed skill path may be read-only -- copy to

  /tmp/skill-name/

  , edit there, package from the copy
• Direct writes may fail; stage in

  /tmp/

  first

Cowork-specific instructions

• Subagents work. If severe timeout problems arise, run test prompts in series.
• No browser/display: use

  --static <output_path>

  for the eval viewer, then provide the link for the user to open locally.
• Generate the eval viewer before evaluating inputs. See Step 4 above.
• Feedback: "Submit All Reviews" downloads

  feedback.json

  . Read from there.
• Description optimization (

  run_loop.py

  ) works fine -- uses

  claude -p

  .
• Save description optimization until the skill is fully done and the user agrees.
• Updating an existing skill: same instructions as Claude.ai section above.

Track these lifecycle steps in TodoWrite if available. In Cowork, specifically add "Create evals JSON and run

eval-viewer/generate_review.py