skill-craft

Create, optimize, or update AgentSkills with evidence-based design. Two modes: (1) CREATE — build new skills from scratch with proven design patterns. (2) OPTIMIZE — diagnose and fix existing skills across 9 evidence-based dimensions: trigger precision, context efficiency, structured instructions, step decomposition, positive constraints, example design, verification, freedom calibration, progressive loading. Based on 25+ peer-reviewed papers (ACL/NeurIPS/EMNLP 2023-2025), Anthropic/Google/Microsoft official guides, and Claude Code/Cursor community practices. Trigger: "create skill", "optimize skill", "improve skill", "fix skill", "package skill", "skill not working", "skill design", "SKILL.md", "写技能", "优化技能", "skill type", "skill classification", "state management", "技能分类"

Build and optimize AI agent skills using evidence-based design principles.

Two Modes of Operation

CREATE mode — Build a new skill from scratch (Steps 1-6). OPTIMIZE mode — Diagnose and fix an existing skill (Step 7).

For new skills, follow Steps 1-6 (Step 7 baked into Step 4). For existing skills needing improvement, jump to Step 7.

---

Skill Classification

Before creating or optimizing, classify the skill type — this determines design priorities.

Nine Skill Types

Type	Code	Definition	Design Priority
Process Control	PROC	Sequential, phased workflows (brainstorming, planning)	D4 Steps → D7 Verify → D2 Efficiency
Methodology	METH	Reusable methods with iron laws (TDD, debugging)	D4 Steps → D7 Verify → D5 Positive
Tool Integration	TOOL	API/CLI wrappers (web-access, calculator)	D1 Description → D8 Freedom → D6 Examples
Domain Knowledge	KNOW	Domain expertise and standards (design, mobile)	D9 Progressive → D2 Efficiency → D3 Structure
Creative/Generative	GEN	Creative output with quality bars (prompt refinement)	D5 Positive → D8 Freedom → D6 Examples
Document Generation	DOC	Structured document output (reports, wikis)	D4 Steps → D3 Structure → D7 Verify
Orchestration	ORCH	Multi-agent/task coordination (parallel dispatch)	D4 Steps → D7 Verify → D6 Examples
Meta-Skill	META	Skills about skills (this very skill)	D3 Structure → D9 Progressive → D5 Positive
Quality Assurance	QA	Verification and review (code review, testing)	D7 Verify → D5 Positive → D4 Steps

See references/skill-taxonomy.md for per-type design templates, pitfalls, and the full classification decision tree.

---

Skill Anatomy

text

skill-name/
├── SKILL.md          (required) — YAML frontmatter + Markdown instructions
├── scripts/          (optional) — Deterministic code (Python/Bash)
├── references/       (optional) — Detailed docs loaded on demand
└── assets/           (optional) — Templates/images used in output

Three-Level Loading (Progressive Disclosure)

Level	Content	When Loaded	Size Target
1	name + description (frontmatter)	Always in context	~100 words
2	SKILL.md body	When skill triggers	< 500 lines, < 5k words
3	scripts/, references/, assets/	On demand	Unlimited

Why this matters: Context window is the scarcest resource. Claude Code Issue #2544 (39👍) proves rules get ignored when SKILL.md is too long. CFPO research (Fudan/Microsoft, 2025) shows format+content joint optimization yields +5~38% improvement.

What NOT to Include

README.md, CHANGELOG.md, installation guides, user-facing docs. Only what the AI agent needs to do the job.

State Management in Skills

Some skills need state management — tracking progress across steps or sessions. Based on OpenSpec's DAG pattern and LangGraph's checkpoint system (see references/state-machine-patterns.md).

Does this skill need state management?

Condition	Needs State?	Pattern
Multi-step workflow with dependencies between steps	YES	DAG + file existence checks
Output of one step feeds into another	YES	Checkpoint pattern
Each invocation is independent	NO	Stateless
Free-form exploration or conversation	NO	Stateless

For state patterns (file existence, checkbox, YAML, DAG) and script integration, see references/state-machine-patterns.md.

Key principle: Use deterministic checks (file exists, script output) rather than LLM judgment for state transitions. Research (arXiv:2511.07585) shows large models have only 12.5% output consistency — scripts are 8× more reliable.

---

CREATE Mode: Steps 1-6

Skill Naming

• Lowercase letters, digits, hyphens only. Under 64 chars.
• Prefer short, verb-led: pdf-editor, db-migrate, brand-styling.
• Namespace by tool when it helps triggering: gh-address-comments.

Step 1: Understand with Concrete Examples

Ask the user (one question at a time):

1. "What should this skill do?"
2. "Can you give 2-3 examples of how it would be used?"
3. "What would a user say that should trigger this skill?"
4. "Are there edge cases or error scenarios to handle?"

Step 2: Plan Resources

For each example from Step 1, identify reusable components:

If the task...	Add this resource
Rewrites the same code each time	scripts/transform.py
Requires domain schemas/docs	references/schema.md
Uses templates or boilerplate	assets/template/

Step 3: Initialize

bash

scripts/init_skill.py <skill-name> --path <output-dir> [--resources scripts,references,assets] [--examples]

Step 4: Write the Skill (with Built-in Optimization)

This step creates SKILL.md following all 9 optimization dimensions from the start. Write using imperative form.

4a. Write Frontmatter (D1: Trigger Precision)

The description field is the only content always loaded. It determines whether the skill triggers at all. Issue #9716 (69👍) documents skills ignored because descriptions lacked trigger phrases.

Three-part formula:

yaml

# BAD — vague, no triggers
description: Helps with PDF operations

# GOOD — capabilities + trigger scenarios + keywords
description: >
  Create, edit, and analyze PDF documents with form filling,
  text extraction, and page manipulation. Use when: (1) Creating PDFs,
  (2) Extracting text or form data, (3) Rotating/merging pages,
  (4) Converting PDFs to images. Trigger: "pdf", "rotate pdf",
  "merge pdf", "extract text", "fill form"

Only name and description fields allowed in frontmatter.

4b. Write Body Sections (D3: Structure)

Separate content types with Markdown headings:

markdown

## Workflow
[Numbered steps — see D4 below]

## Constraints
[Specific, measurable rules — see D5 below]

## Examples
[Concrete user request → agent behavior — see D6 below]

## Verification
[Success criteria and how to check — see D7 below]

## Resources
[Links to references/ — see D9 below]

4c. Write Workflows (D4: Step Decomposition)

Research (CoM, ACL 2025; Watch Every Step, EMNLP 2024) proves numbered steps with decision points outperform free-form descriptions.

markdown

# BAD — paragraph description
Process the document by extracting key fields and validating them
against the schema, then generating a report.

# GOOD — numbered steps with branching
## Workflow

1. **Extract**: Read document, extract fields in `references/schema.md`
2. **Validate**: Check each field against schema rules
   - If valid → proceed to step 3
   - If invalid → report specific errors, suggest fixes, STOP
3. **Transform**: Run `scripts/transform.py --input <extracted.json>`
4. **Report**: Generate output using `assets/report-template.md`
5. **Verify**: Confirm output contains all required sections

4d. Write Constraints (D5: Positive Framing)

Inverse IFEval (ByteDance, 2025, 1012 test cases) proves positive constraints are followed significantly more reliably. Constraint Decomposition (2025): 41.2% → 73.8% accuracy when constraints are decomposed.

markdown

# BAD — negative constraints
Don't write functions longer than 20 lines.
Never modify files outside the project.
Avoid using deprecated APIs.

# GOOD — positive constraints with specifics
Keep every function under 20 lines. Extract helpers for complex logic.
Only modify files within the target directory specified by the user.
Use APIs from the approved list in `references/api-versions.md`.

For safety-critical rules, keep negative framing but pair with positive alternative:

markdown

CRITICAL: Only delete files after explicit user confirmation. (safety negative)
Prefer creating new files over modifying existing ones when possible. (positive alternative)

4e. Write Examples (D6: Example Design)

Cursor official data: Good/Bad examples make rules ~3x more effective. ACL 2025: 1-shot achieves best flexibility/structure balance.

markdown

# BAD — abstract
Use the transform script to convert between formats.

# GOOD — concrete "user request → agent behavior"
## Examples

### User: "Convert this CSV to JSON"
→ 1. Run `scripts/transform.py data.csv --output data.json`
→ 2. Verify JSON has keys: `name`, `date`, `amount`
→ 3. Report: "Converted 150 records from CSV to JSON"

### User: "The output has wrong date format"
→ 1. Check current date format in output
→ 2. Add `--date-format ISO8601` flag
→ 3. Re-run and verify dates match `YYYY-MM-DD`

Include 1-3 examples in SKILL.md. For more, create references/examples.md.

4f. Write Verification (D7: Verification)

Issue #42796 (1781👍): Analysis of 6,852 sessions shows agent Read:Edit ratio dropped 70%, causing "rush to completion" errors. Explicit verification steps prevent this.

markdown

## Verification

After completing the workflow:
1. **Output exists**: Confirm the output file was created
2. **Schema valid**: Run `scripts/validate.py output.json`
3. **Content complete**: Output has all required sections
4. **No regressions**: Run `npm test`

If any check fails → diagnose, fix, re-verify.

4g. Calibrate Freedom (D8: Freedom Calibration)

Match specificity to task fragility (NeurIPS 2024: task type determines optimal approach):

Task Fragility	Freedom Level	How to Write
Fragile (deployment, DB migration, security)	Low — specific scripts + numbered steps	"Run scripts/backup.sh then scripts/migrate.py --dry-run"
Moderate (API calls, code generation)	Medium — preferred patterns + alternatives	"Prefer pdfplumber. For encrypted PDFs, use PyMuPDF instead."
Open (writing, brainstorming, design)	High — principles + examples	"Match brand voice in references/brand-voice.md. Target 500-800 words."

4h. Manage Context (D2 + D9: Efficiency + Progressive Loading)

D2 — Token economy: Every line must earn its cost. Delete what the model already knows. One example beats three paragraphs.

D9 — Progressive loading: Keep SKILL.md < 500 lines. Move details to references/ with explicit "when to read" pointers:

markdown

## Resources

- **Advanced schemas**: See [SCHEMAS.md](references/schemas.md) when user provides non-standard input
- **Error handling**: See [TROUBLESHOOTING.md](references/troubleshooting.md) when scripts fail
- **API reference**: See [API.md](references/api.md) when working with specific endpoints

Keep references one level deep (no nested references). Files > 100 lines: include table of contents at top.

Step 5: Package

bash

scripts/package_skill.py <path/to/skill-folder> [output-dir]

Automatically validates structure, frontmatter, naming, and description quality before packaging.

Step 6: Iterate

1. Use the skill on real tasks
2. Notice struggles or inefficiencies
3. Apply Step 7 (Optimize) to systematically improve

---

OPTIMIZE Mode: Step 7

Use when an existing skill needs improvement. Read references/skill-optimization.md for detailed before/after examples and references/evidence-database.md for research sources.

7.1 Diagnose

Read the full skill. Answer each diagnostic question:

ID	Dimension	Diagnostic Question	Impact if Failing
D1	Description	Seeing ONLY name+description, would an AI trigger this skill correctly?	Skill never triggers or triggers incorrectly
D2	Efficiency	Does every line earn its token cost?	Context bloat degrades all loaded skills
D3	Structure	Are workflows/constraints/examples visually separated?	Instructions blur together, compliance drops
D4	Steps	Are complex processes numbered with decision points?	Agent improvises, misses critical sequence
D5	Positive	Are rules "do X" not "don't do Y"?	Negative constraints poorly followed (IFEval data)
D6	Examples	Is there a concrete "user request → behavior" example?	Agent guesses expected behavior
D7	Verification	Does the skill define how to check output?	Agent "rushes to completion" (Issue #42796)
D8	Freedom	Is specificity matched to task fragility?	Over/under-constrained output
D9	Progressive	Is SKILL.md lean with details in references/?	All content loaded even when not needed

Run scripts/quick_validate.py <skill-path> for automated hints on D1, D2, D5, D6, D7, D9.

7.2 Prioritize by Skill Type

Focus on the dimensions that matter most for this skill type:

Skill Type	Fix These First	Why
Workflow (sequential process)	D4 Steps → D7 Verify → D2 Efficiency	Sequence errors are catastrophic; verification catches them
Tool Integration (API/CLI wrapper)	D1 Description → D8 Freedom → D6 Examples	Must trigger on right tool names; right specificity level
Knowledge/Reference (schemas, policies)	D9 Progressive → D2 Efficiency → D3 Structure	Massive content needs aggressive context management
Creative/Generative (writing, design)	D5 Positive → D8 Freedom → D6 Examples	Needs principles not rigid rules; examples set quality bar
Debugging/Analysis (troubleshooting)	D4 Steps → D7 Verify → D5 Positive	Must follow diagnostic sequence; verify fix actually works
Document Generation (reports, analysis)	D4 Steps → D3 Structure → D7 Verify	Output structure must be fixed; verification ensures completeness
Orchestration (multi-agent, parallel tasks)	D4 Steps → D7 Verify → D6 Examples	Dispatch logic must be precise; sub-agent prompts need templates
Meta-Skill (skills about skills)	D3 Structure → D9 Progressive → D5 Positive	Complex logic needs clean structure; must avoid recursive bloat
Quality Assurance (review, testing, verification)	D7 Verify → D5 Positive → D4 Steps	Verification is the core function; positive framing prevents over-flagging

7.3 Apply Dimension Fixes

For each failing dimension, apply the fix. See references/skill-optimization.md for detailed before/after examples.

D1 — Rewrite description with 3-part formula: (1) capabilities, (2) trigger scenarios, (3) keywords.

D2 — Audit every line: delete known knowledge, compress with examples, move details to references/. Target < 500 lines.

D3 — Add visual hierarchy: Markdown ##/### headings for each content type. --- between major blocks.

D4 — Number the steps: Convert paragraphs to numbered lists. Add if/else decision points. Add error handling.

D5 — Flip to positive: Convert "don't X" → "do Y". For safety rules, keep negative but add positive alternative.

D6 — Add examples: 1-3 concrete "User: ... → Agent behavior: ..." examples. Realistic inputs, not placeholders.

D7 — Add verification: Explicit success criteria at workflow end. What to check, how to check, what if it fails.

D8 — Calibrate freedom: Fragile tasks → numbered steps/scripts. Creative tasks → principles/examples.

D9 — Move to references: Anything beyond core workflow → references/. Each link says WHEN to read it.

7.4 Scan for Anti-Patterns

Anti-Pattern	Symptom	Fix
Kitchen Sink	SKILL.md > 500 lines	Split into focused skills or references/
Invisible Skill	No trigger phrases in description	Rewrite with Trigger: keywords
Vague Guardrails	"Write clean code", "Be helpful"	Replace with specific, measurable rules
Negative-Only	Mostly "don't" / "never" rules	Convert to positive framing
Missing Exit	No success/failure criteria	Add verification section
Flat Wall of Text	No headings or visual breaks	Add ## / ### headings
Zombie References	Files exist but unlinked from SKILL.md	Add explicit "See X when Y" pointers
Example-Free	No concrete examples	Add 1-3 user request → behavior examples
Stateless Multi-Step	Multi-step workflow with no state tracking	Add file existence checks or checkbox tracking
LLM-as-State-Oracle	Using LLM to determine workflow state instead of scripts	Replace with deterministic checks (file exists, script output)
Missing Script Integration	References to scripts/ but no scripts/ directory	Create the scripts or remove references

7.5 Validate After Optimization

1. Run scripts/quick_validate.py <skill-path> — structure + optimization hints
2. Re-read end-to-end — does each section serve a clear purpose?
3. Count lines — SKILL.md body still < 500?
4. Check all reference links — valid? Explain when to read?

---

Reference Files

File	When to Read
references/skill-optimization.md	Optimizing a skill — detailed before/after examples for all 9 dimensions
references/evidence-database.md	Checking research evidence — papers, data, community issues
references/skill-taxonomy.md	Classifying skill types — 9 types, design templates, user need routing
references/state-machine-patterns.md	Adding state management — DAG patterns, checkpoint design, script-driven workflows

Scripts

Script	Purpose
scripts/init_skill.py	Create new skill directory with optimized template
scripts/package_skill.py	Validate + package skill into .skill file
scripts/quick_validate.py	Validate structure + check optimization hints