Skill Creator

Skill Creator

Create skills and iteratively improve them through measurement.

Output style directive (applies to every generated skill/agent)

Generated SKILL.md and agent bodies must be written as dense informational text focused on accuracy. Minimize prose, maximize signal, no filler.

• No motivational framing, no pep talk, no "this will help you...".
• No repeated restatement of the same constraint in different words.
• Prefer tables, numbered phases, and bullet lists over paragraphs.
• Every sentence must carry information the model will act on; cut anything that is only atmosphere.
• Explanations of "why" stay short and attached to the rule they justify -- one clause, not a paragraph.

New and edited skill files are written to the Dense-Complete Writing standard: skills/shared-patterns/dense-complete-writing.md. That file is the canonical rule; the bullets above are its application to skill scaffolding.

This is a generation constraint on the outputs of this skill, not a style note for this skill's own prose. Enforce it during the "Write the SKILL.md" phase and during any agent scaffolding.

The process:

• Decide what the skill should do and how it should work
• Write a draft of the skill
• Create test prompts and run claude-with-the-skill on them
• Evaluate the results -- both with agent reviewers and optionally human review
• Improve the skill based on what the evaluation reveals
• Repeat until the skill demonstrably helps

Figure out where the user is in this process and help them progress. If they say "I want to make a skill for X", help narrow scope, write a draft, write test cases, and run the eval loop. If they already have a draft, go straight to testing.

Creating a skill

Capture intent

Start by understanding what the user wants. The current conversation might already contain a workflow worth capturing ("turn this into a skill"). If so, extract:

1. What should this skill enable Claude to do?
2. When should this skill trigger? (what user phrases, what contexts)
3. What is the expected output?
4. Are the outputs objectively verifiable (code, data transforms, structured files) or subjective (writing quality, design aesthetics)? Objectively verifiable outputs benefit from test cases. Subjective outputs are better evaluated by human review.

Duplicate Domain Check

Before creating any new skill, check whether an existing umbrella skill already covers this domain. This is mandatory -- skipping it leads to system prompt bloat and routing degradation.

Step 1: Search for existing domain coverage.

grep -i "<domain-keyword>" skills/INDEX.json
ls skills/ | grep "<domain-prefix>"

Step 2: If a domain skill exists, determine whether the new skill's scope is a sub-concern of the existing skill. Sub-concerns MUST be added as reference files on the existing skill, not created as separate skills.

Pattern (correct): skills/perses/references/plugins.md Incorrect (creates domain overlap): skills/perses-plugin-creator/SKILL.md

Step 3: If no domain skill exists and the domain has multiple sub-concerns, create the skill with a references/ directory from the start.

One domain = one skill + many reference files. Never create multiple skills for the same domain.

Only proceed to writing a new SKILL.md if no existing skill covers the domain, or if the user explicitly confirms creating a new skill after reviewing the overlap.

Research

Read docs/PHILOSOPHY.md before writing any component. The philosophy contains binding architectural decisions — not suggestions — that govern how agents carry knowledge, how skills structure workflows, how references are organized, and how content is framed. Components that violate the philosophy will fail review.

Read the repository CLAUDE.md before writing anything. Project conventions override default patterns.

Write the SKILL.md

Based on the user interview, create the skill directory and write the SKILL.md.

Skill structure:

skill-name/
├── SKILL.md              # Required -- the workflow
├── SPEC.md               # Optional -- contract for complex/high-impact skills
├── EVAL.md               # Optional -- repeatable eval cases for complex/high-impact skills
├── scripts/              # Deterministic CLI tools the skill invokes
├── agents/               # Subagent prompts used only by this skill
├── references/           # Deep context loaded on demand
└── assets/               # Templates, viewers, static files

Maintenance artifacts -- For Complex skills, security-sensitive skills, router-facing skills, PR/release workflows, and skills likely to be iterated over time, create SPEC.md and EVAL.md alongside SKILL.md:

• SPEC.md: purpose, scope, non-goals, invariants, dependencies, and success criteria. It is the contract maintainers use when changing the skill.
• EVAL.md: representative prompts/cases, expected routing or behavior, known failure modes, and pass/fail checks. It is the regression suite for skill behavior.

Do not create SOURCES.md as a standard artifact. Provenance belongs in docs, ADRs, citations, or research files when it matters. If the LLM does not need the file to execute, evaluate, or maintain the component, keep it out of the component directory.

Maintenance artifacts are not runtime context. SKILL.md should not instruct the model to load SPEC.md or EVAL.md during ordinary execution. Load them only when creating, evaluating, redesigning, or modifying the skill.

Frontmatter -- name, description, routing metadata. Description caps: 60 chars max for non-invocable skills, 120 chars for user-invocable. No "Use when:", "Use for:", or "Example:" in the description. The /do router has its own routing tables.

user_invocable default is false. New skills are agent-facing by default: the /do router dispatches them, and the user never types the skill name. Emit the frontmatter field explicitly so the default is visible:

user_invocable: false  # default -- router-dispatched, not user-typed

Flipping to true requires an explicit justification comment in the frontmatter naming the user-facing trigger phrases and why routing through /do is insufficient. Example:

user_invocable: true  # justification: users type "/pr-workflow" directly as
                      # a slash-command entry point; /do dispatch is bypassed
                      # because the user is already scoped to the PR lifecycle.

No justification = leave it false. User-invocable expands the system-prompt surface and the slash-command namespace; both are scarce.

force_route -- when to set true. Adds force_bonus = 2.0 to the router match score (see force_bonus in scripts/pre-route.py). Without it, single-trigger skills cap at "low" confidence and may not route at all -- the bug PR #663 fixed for agent-creator. Confidence ladder: force_route + 2+ triggers = high; force_route + 1 trigger = medium; 3+ triggers without force = medium; otherwise low.

Set force_route: true when the skill matches one of these patterns:

Leave force_route off and earn confidence through trigger count when the skill matches one of these patterns:

Rule of thumb: if a future contributor reading the trigger list could reasonably ask "did the user mean this or that?", leave force_route off and rely on multiple triggers for confidence.

See references/skill-template.md for the complete frontmatter template with all fields and valid values.

Frontmatter validation (mandatory post-write gate): After writing SKILL.md, validate YAML frontmatter:

python3 scripts/validate-skill-frontmatter.py skills/<skill-name>/SKILL.md

Scaffold is not complete until this exits 0. The validator catches: broken YAML, name/directory mismatch, missing routing section, missing triggers, missing category, top-level pairs_with, and force_routing typo.

The description is the primary triggering mechanism. Claude tends to undertrigger skills -- be explicit about trigger contexts. Include "Use for" with concrete phrases users would say.

Body -- workflow first, then context:

1. Brief overview (2-3 sentences: what this does and how)
2. Instructions / workflow phases (the actual methodology)
3. Reference material (commands, guides, schemas)
4. Error handling (cause/solution pairs for common failures)
5. References to bundled files

Constraints belong inline within the workflow step where they apply. Explain the reasoning behind constraints -- "Run with -race because race conditions are silent until production" generalizes; "ALWAYS run with -race" does not.

Do-pair validation -- After writing any failure-mode blocks, run:

python3 scripts/validate-references.py --check-do-framing

Every failure-mode block must have a paired "Do instead" counterpart. Blocks without one fail the check. If a prohibition genuinely has no correct alternative, annotate it with <!-- no-pair-required: reason --> to pass validation without a "Do instead" block. Ship the skill only after this check exits 0.

Triple-validation verdicts on documented patterns -- When the skill documents patterns the model is supposed to apply (mental models, heuristics, phrase fingerprints, voice traits, code conventions), every pattern block carries an explicit verdict from the triple-validation rubric: KEEP, FOOTNOTE, or DROP. KEEP and FOOTNOTE patterns ship in the SKILL.md; DROP patterns stay in working notes (pattern-candidates.md or equivalent) and never reach the published file.

The rubric (recurrence, generative power, exclusivity) lives at skills/content/create-voice/references/extraction-validation.md. Load it on demand when running the gate; do not duplicate the content here.

Three accepted verdict markers, in priority order:

1. A line in the pattern's body containing **Verdict**: KEEP / **Verdict**: FOOTNOTE / **Verdict**: DROP.
2. An inline tag at the end of the H3 heading: ### M1: Mechanism-first (KEEP).
3. A blanket tag on the parent H2 heading covering every child: ## Mental Models (KEEP-verdict). Per-block markers override the blanket.

For each KEEP or FOOTNOTE pattern, attach one line of evidence covering the three checks ("appears in X and Y; predicts Z; distinguishes from peer W"). Patterns without that evidence fail the gate even if they carry a verdict word -- the verdict is a claim, the evidence is what makes it auditable.

Phase gate (before shipping): every documented pattern carries a KEEP or FOOTNOTE verdict with one-line evidence covering the three checks. Patterns without verdicts fail the gate. The deterministic check below enforces the verdict half; the evidence half is read by reviewers.

python3 scripts/check-skill-verdicts.py skills/<your-skill>/SKILL.md

The script walks H3 sections under H2 parents named Mental Models, Heuristics, Phrase Fingerprints, or Patterns (case-insensitive substring match) and exits non-zero on any block lacking a KEEP or FOOTNOTE verdict, or carrying DROP. Wire it into the post-scaffold gate alongside validate-references.py. Skills that document no patterns (pure workflow skills) exit 0 trivially -- the gate only fires when there are patterns to verdict.

Progressive disclosure -- SKILL.md is the routing target, not the reference library. It stays lean so it loads fast when Claude considers invoking it, then reads references/ on demand as phases execute. See references/progressive-disclosure.md for the full model, economics, and extraction decision tree.

Key rules:

• SKILL.md: brief overview, phase structure with gates, one-line pointers to reference files, error handling
• references/: checklists, rubrics, agent dispatch prompts, report templates, pattern catalogs, example collections -- anything only needed at execution time
• If SKILL.md exceeds 500 lines after writing, extract detailed content to references/ before proceeding
• If SKILL.md exceeds 700 lines, extraction is mandatory -- it is carrying reference content that should not be loaded on every routing decision

The most effective complex skills (sapcc-review, voice-writer) keep SKILL.md under 600 lines and put operational depth in references/ and agents/. Rich references/ content adds depth at zero routing cost; deterministic scripts/ ensure consistency; bundled agents/ prompts enable specialized dispatch without routing overhead.

See references/progressive-disclosure.md for the real numbers and extraction decision tree.

Bundled scripts

Extract deterministic, repeatable operations into scripts/*.py CLI tools with argparse interfaces. Scripts save tokens (the model doesn't reinvent the wheel each invocation), ensure consistency across runs, and can be tested independently.

Pattern: scripts/ for deterministic ops, SKILL.md for LLM-orchestrated workflow.

Bundled agents

For skills that spawn subagents with specialized roles, bundle agent prompts in agents/. These are not registered in the routing system -- they are internal to the skill's workflow.

Agent creation standard

When this skill scaffolds a repo-level agent, read docs/PHILOSOPHY.md first. The philosophy governs how agents carry knowledge (not thin wrappers), how review knowledge separates from implementation knowledge, and how references are organized for progressive disclosure. An agent built without reading the philosophy will misplace domain knowledge or violate structural conventions.

Apply the same maintenance-artifact rule to the agent package:

agents/
├── {agent-name}.md
└── {agent-name}/
    ├── SPEC.md            # Optional -- contract for complex/high-impact agents
    ├── EVAL.md            # Optional -- repeatable eval cases
    └── references/
        └── ...

Use SPEC.md and EVAL.md for agents that are complex, high-impact, security-sensitive, router-facing, or likely to be tuned repeatedly. Do not create SOURCES.md as a default agent artifact.

Path placeholder convention (per ADR-201)

Author-machine paths leak into reference docs as casually as paste-from-shell-history. They look like stable interfaces ("see /tmp/foo/ for the canonical layout") but they exist on the author's box only. A user who follows the doc literally arrives at a path that does not exist — or, worse, finds an unrelated /tmp/foo/ from someone else's tooling.

Any path appearing in a published reference doc that matches one of the following MUST be either an angle-bracket placeholder OR a path explicitly labelled as the author's local validation harness:

• /tmp/...
• /home/<author>/...
• ~/<author-specific>/... (anything with a user-name segment)
• /Users/<author>/... (macOS user dir)
• /private/var/folders/... (macOS tmp dir)

Form (a) — placeholder (preferred):

Place outputs at `<your-output-dir>/assets/<slug>/final.png`.

Form (b) — labelled author-harness (only when the path has documentary value as a worked example):

> **Author's local validation harness, replace before use:** `/tmp/sprite-demo/...`

Form (b) is allowed only when the path is referenced for evidence (e.g. "in our test run we observed X at /tmp/sprite-demo/foo.png"). Form (a) is preferred everywhere else.

Integration-target paths stay as-is. When a skill explicitly knows about a target project — ~/road-to-aew, ~/deeproute — those references stay literal. The skill's frontmatter and body explain that the skill targets that project; the path is part of the integration contract, not a leak.

Authoring-time enforcement. Before declaring a skill shippable, run the toolkit-wide audit grep. Any non-empty result is a violation requiring placeholder replacement:

grep -rnE '(/tmp/[a-z][a-z0-9_-]+|/home/[a-z][a-z0-9_-]+|/Users/[a-z][a-z0-9_-]+)' \
  skills/<your-skill>/SKILL.md skills/<your-skill>/references/*.md 2>/dev/null \
  | grep -v '<your-' \
  | grep -v "Author's local validation harness"

The validation pass invoked at the post-scaffold gate (next section) includes this grep. New skills do not ship with leaked author paths.

Post-scaffold: regenerate skills INDEX.json (mandatory)

After the skill directory + SKILL.md are on disk, regenerate the skills index. Without this step the router cannot discover the new skill and requests that should match it fall through to the fallback handler.

python3 scripts/generate-skill-index.py

Run it from the repo root. Treat it as a commit-gating step: the scaffold is not complete until INDEX.json reflects the new skill. Diff the file before staging to confirm exactly one new entry was added.

Post-scaffold: content-cleanliness audit

Audit the SKILL.md for non-runtime content (meta-commentary, changelog prose, narration the runtime never reads). The audit globs <root>/*/SKILL.md, so point --root at the new skill's category directory to scan it:

python3 scripts/audit-skill-content.py --root skills/<category> --severity high

Exit 0 with zero high-severity violations is the gate. Drop to --severity low to see every flagged line. This catches content that bloats the skill without serving the runtime — distinct from condense (prose density) and joy-check (framing).

Post-scaffold: joy-check + do-pair validation

Before declaring the skill shippable, run both checks. They catch different failure modes: joy-check catches grievance-mode framing that drags the model toward pessimism; do-pair validation catches failure modes with no paired "Do instead" counterpart.

Joy-check (framing). Invoke the joy-check skill on the SKILL.md and each references/*.md file. The accepted deterministic substitute is:

python3 scripts/validate-references.py --check-do-framing

This script enforces the positive-pairing rule that joy-check encodes structurally: every failure mode gets a constructive counterpart. Use it when dispatching the full joy-check skill is disproportionate (small edits, CI contexts, or when only structural pairing matters). For any new skill that ships prose-heavy references, prefer the full joy-check skill -- tone drift is not caught by the pairing script.

Do-pair validation (structural). Same command, different failure class. Ship the skill only after this exits 0:

python3 scripts/validate-references.py --check-do-framing

Post-scaffold: condense

After creation and validation, run the condense skill on the new SKILL.md to maximize information density. The condense skill strips prose filler while preserving every instruction, rule, gate, and code block — because new skills tend to ship verbose and the condense pass catches what the author's eye skips.

Testing the skill

This is the core of the eval loop. Do not stop after writing -- test the skill against real prompts and measure whether it actually helps.

Create test prompts

Write 2-3 realistic test prompts -- the kind of thing a real user would say. Rich, detailed, specific. Not abstract one-liners.

Bad: "Format this data" Good: "I have a CSV in ~/downloads/q4-sales.csv with revenue in column C and costs in column D. Add a profit margin percentage column and highlight rows where margin is below 10%."

Share prompts with the user for review before running them.

See references/bundled-components.md for the evals.json format and workspace directory layout.

Run test prompts

For each test case, spawn two subagents in the same turn -- one with the skill loaded, one without (baseline). Launch everything at once so it finishes together.

With-skill run: Tell the subagent to read the skill's SKILL.md first, then execute the task. Save outputs to the workspace.

Baseline run: Same prompt, no skill loaded. Save to a separate directory.

Evaluate results

Evaluation has three tiers, applied in order:

Tier 1: Deterministic checks -- run automatically where applicable:

• Does the code compile? (go build, tsc --noEmit, python -m py_compile)
• Do tests pass? (go test -race, pytest, vitest)
• Does the linter pass? (go vet, ruff, biome)

Tier 2: Agent blind review -- dispatch using agents/comparator.md:

• Comparator receives both outputs labeled "Output 1" / "Output 2"
• It does NOT know which is the skill version
• Scores on relevant dimensions, picks a winner with reasoning
• Save results to blind_comparison.json

Tier 3: Human review (optional) -- generate the comparison viewer:

python3 scripts/eval_compare.py path/to/workspace
open path/to/workspace/compare_report.html

The viewer shows outputs side by side with blind labels, agent review panels, deterministic check results, winner picker, feedback textarea, and a skip-to-results option. Human reviews are optional -- agent reviews are sufficient for iteration.

Draft assertions

While test runs are in progress, draft quantitative assertions for objective criteria. Good assertions are discriminating -- they fail when the skill doesn't help and pass when it does. Non-discriminating assertions ("file exists") provide false confidence.

Run the grader (agents/grader.md) to evaluate assertions against outputs:

• PASS requires genuine substance, not surface compliance
• The grader also critiques the assertions themselves -- flagging ones that would pass regardless of skill quality

Aggregate results with scripts/aggregate_benchmark.py to get pass rates, timing, and token usage with mean/stddev across runs.

Improving the skill

This is the iterative heart of the process.

Generalize from feedback. If a fix only helps the test case but wouldn't generalize, it's overfitting. Try different approaches rather than fiddly adjustments.

Keep instructions lean. Read execution transcripts, not just final outputs. Remove instructions that cause the model to waste time -- they consume attention budget without producing value.

Explain the reasoning. Motivation-based instructions generalize better than bare imperatives. "Prefer X because Y" lets the model apply the principle to situations the skill author didn't anticipate.

Extract repeated work. If all subagents independently wrote similar helper scripts, bundle that script in scripts/. One shared implementation beats N independent reinventions.

The iteration loop

1. Apply improvements to the skill
2. Rerun all test cases into iteration-<N+1>/, including baselines
3. Generate the comparison viewer with --previous-workspace pointing at the prior iteration
4. Review -- agent or human
5. Repeat until results plateau or the user is satisfied

Stop iterating when:

• Feedback is empty (outputs look good)
• Pass rates aren't improving between iterations
• The user says they're satisfied

Description optimization

After the skill works well, optimize the description for triggering accuracy.

See references/bundled-components.md for the full optimization loop: eval query format, train/test split, optimize_description.py usage, and overfitting guards.

Enriching existing skills

Use this mode when a skill already exists but produces shallow, generic output -- it has thin references/, no scripts/, and passes an eval by luck rather than by containing domain knowledge that changes behavior.

Indicators this mode is appropriate:

• references/ has fewer than 2 files, or none at all
• No scripts/ directory
• Eval outputs look plausible but lack domain idioms, concrete examples, or checklists specific to the skill's domain
• The skill passes a test because the model already knows the domain, not because the skill contributes anything

The enrichment loop

Six phases: AUDIT (measure current depth), RESEARCH (find gaps), ENRICH (add reference content), TEST (A/B vs baseline), EVALUATE (blind comparator), PUBLISH (branch + PR). Max 3 iterations before escalating to the user. Each retry uses a different research angle: iteration 1 = official docs, iteration 2 = common mistakes, iteration 3 = advanced patterns.

See references/enrichment-workflow.md for the full phase-by-phase checklist, scoring details, retry logic, and exact commit/PR flow.

Bundled agents and scripts

See references/bundled-components.md for the full list of bundled agents (grader.md, comparator.md, analyzer.md), bundled scripts, workspace layout, and evals.json format.

Reference files

• references/progressive-disclosure.md -- The disclosure model: economics, size gates, what to extract, real examples from the toolkit, script and agent patterns
• references/skill-template.md -- Complete SKILL.md template with all sections
• references/artifact-schemas.md -- JSON schemas for eval artifacts (evals.json, grading.json, benchmark.json, comparison.json, timing.json, metrics.json)
• references/complexity-tiers.md -- Skill examples by complexity tier
• references/workflow-patterns.md -- Reusable phase structures and gate patterns
• references/error-catalog.md -- Common skill creation errors with solutions
• references/enrichment-workflow.md -- Deep reference for the enrichment loop: AUDIT checklist, RESEARCH strategy, ENRICH structuring, TEST/EVALUATE/PUBLISH phases, and retry logic in detail
• references/domain-research-targets.md -- Lookup table: given a skill's domain, which primary sources, secondary sources, and extraction targets to use during RESEARCH
• references/bundled-components.md -- Bundled agents, scripts, workspace layout, evals.json format, and description optimization procedure

Error handling

Skill doesn't trigger when it should

Cause: Description is too vague or missing trigger phrases Solution: Add explicit "Use for" phrases matching what users actually say. Test with scripts/optimize_description.py.

Test run produces empty output

Cause: The claude -p subprocess didn't load the skill, or the skill path is wrong Solution: Verify the skill directory contains SKILL.md (exact case). Check the --skill-path argument points to the directory, not the file.

Grading results show all-pass regardless of skill

Cause: Assertions are non-discriminating (e.g., "file exists") Solution: Write assertions that test behavior, not structure. The grader's eval critique section flags these -- read it.

Iteration loop doesn't converge

Cause: Changes are overfitting to test cases rather than improving the skill Solution: Expand the test set with more diverse prompts. Focus improvements on understanding WHY outputs differ, not on patching specific failures.

Description optimization overfits to train set

Cause: Test set is too small or train/test queries are too similar Solution: Ensure should-trigger and should-not-trigger queries are realistic near-misses, not obviously different. The 60/40 split guards against this, but only if the queries are well-designed.

Reference Loading Table