Writing Skills

Overview

Writing skills IS Test-Driven Development applied to process documentation.

Personal skills live in your runtime’s skills directory (~/.claude/skills/ on Claude Code) — see codex-tools.md (opens in a new tab) or gemini-tools.md (opens in a new tab) for the path on those runtimes. Codex, Copilot CLI, and Gemini CLI all also recognize ~/.agents/skills/ as a cross-runtime alias.

You write test cases (pressure scenarios with subagents), watch them fail (baseline behavior), write the skill (documentation), watch tests pass (agents comply), and refactor (close loopholes).

Core principle: If you didn’t watch an agent fail without the skill, you don’t know if the skill teaches the right thing.

REQUIRED BACKGROUND: You MUST understand superpowers:test-driven-development before using this skill. That skill defines the fundamental RED-GREEN-REFACTOR cycle. This skill adapts TDD to documentation.

Official guidance: For Anthropic’s official skill authoring best practices, see anthropic-best-practices.md. This document provides additional patterns and guidelines that complement the TDD-focused approach in this skill.

What is a Skill?

A skill is a reference guide for proven techniques, patterns, or tools. Skills help future agents find and apply effective approaches.

Skills are: Reusable techniques, patterns, tools, reference guides

Skills are NOT: Narratives about how you solved a problem once

TDD Mapping for Skills

The entire skill creation process follows RED-GREEN-REFACTOR.

When to Create a Skill

Create when:

• Technique wasn’t intuitively obvious to you
• You’d reference this again across projects
• Pattern applies broadly (not project-specific)
• Others would benefit

Don’t create for:

• One-off solutions
• Standard practices well-documented elsewhere
• Project-specific conventions (put in your instructions file)
• Mechanical constraints (if it’s enforceable with regex/validation, automate it—save documentation for judgment calls)

Skill Types

Technique

Concrete method with steps to follow (condition-based-waiting, root-cause-tracing)

Pattern

Way of thinking about problems (flatten-with-flags, test-invariants)

Reference

API docs, syntax guides, tool documentation (office docs)

Directory Structure

skills/
  skill-name/
    SKILL.md              # Main reference (required)
    supporting-file.*     # Only if needed

Flat namespace - all skills in one searchable namespace

Separate files for:

1. Heavy reference (100+ lines) - API docs, comprehensive syntax
2. Reusable tools - Scripts, utilities, templates

Keep inline:

• Principles and concepts
• Code patterns (< 50 lines)
• Everything else

SKILL.md Structure

Frontmatter (YAML):

• Two required fields: name and description (see agentskills.io/specification (opens in a new tab) for all supported fields)
• Max 1024 characters total
• name: Use letters, numbers, and hyphens only (no parentheses, special chars)
• description: Third-person, describes ONLY when to use (NOT what it does)
  • Start with “Use when…” to focus on triggering conditions
  • Include specific symptoms, situations, and contexts
  • NEVER summarize the skill’s process or workflow (see SDO section for why)
  • Keep under 500 characters if possible

---
name: Skill-Name-With-Hyphens
description: Use when [specific triggering conditions and symptoms]
---

# Skill Name

## Overview
What is this? Core principle in 1-2 sentences.

## When to Use
[Small inline flowchart IF decision non-obvious]

Bullet list with SYMPTOMS and use cases
When NOT to use

## Core Pattern (for techniques/patterns)
Before/after code comparison

Skill Discovery Optimization (SDO)

Critical for discovery: Future agents need to FIND your skill

1. Rich Description Field

Purpose: Your agent reads the description to decide which skills to load for a given task. Make it answer: “Should I read this skill right now?”

Format: Start with “Use when…” to focus on triggering conditions

CRITICAL: Description = When to Use, NOT What the Skill Does

The description should ONLY describe triggering conditions. Do NOT summarize the skill’s process or workflow in the description.

Why this matters: Testing revealed that when a description summarizes the skill’s workflow, an agent may follow the description instead of reading the full skill content. A description saying “code review between tasks” caused an agent to do ONE review, even though the skill’s flowchart clearly showed TWO reviews (spec compliance then code quality).

When the description was changed to just “Use when executing implementation plans with independent tasks” (no workflow summary), the agent correctly read the flowchart and followed the two-stage review process.

The trap: Descriptions that summarize workflow create a shortcut agents will take. The skill body becomes documentation agents skip.

# ❌ BAD: Summarizes workflow - agents may follow this instead of reading skill
description: Use when executing plans - dispatches subagent per task with code review between tasks

# ❌ BAD: Too much process detail
description: Use for TDD - write test first, watch it fail, write minimal code, refactor

# ✅ GOOD: Just triggering conditions, no workflow summary
description: Use when executing implementation plans with independent tasks in the current session

# ✅ GOOD: Triggering conditions only
description: Use when implementing any feature or bugfix, before writing implementation code

Content:

• Use concrete triggers, symptoms, and situations that signal this skill applies
• Describe the problem (race conditions, inconsistent behavior) not language-specific symptoms (setTimeout, sleep)
• Keep triggers technology-agnostic unless the skill itself is technology-specific
• If skill is technology-specific, make that explicit in the trigger
• Write in third person (injected into system prompt)
• NEVER summarize the skill’s process or workflow

# ❌ BAD: Too abstract, vague, doesn't include when to use
description: For async testing

# ❌ BAD: First person
description: I can help you with async tests when they're flaky

# ❌ BAD: Mentions technology but skill isn't specific to it
description: Use when tests use setTimeout/sleep and are flaky

# ✅ GOOD: Starts with "Use when", describes problem, no workflow
description: Use when tests have race conditions, timing dependencies, or pass/fail inconsistently

# ✅ GOOD: Technology-specific skill with explicit trigger
description: Use when using React Router and handling authentication redirects

2. Keyword Coverage

Use words an agent would search for:

• Error messages: “Hook timed out”, “ENOTEMPTY”, “race condition”
• Symptoms: “flaky”, “hanging”, “zombie”, “pollution”
• Synonyms: “timeout/hang/freeze”, “cleanup/teardown/afterEach”
• Tools: Actual commands, library names, file types

3. Descriptive Naming

Use active voice, verb-first:

• ✅ creating-skills not skill-creation
• ✅ condition-based-waiting not async-test-helpers

4. Token Efficiency (Critical)

Problem: getting-started and frequently-referenced skills load into EVERY conversation. Every token counts.

Target word counts:

• getting-started workflows: <150 words each
• Frequently-loaded skills: <200 words total
• Other skills: <500 words (still be concise)

Techniques:

Move details to tool help:

# ❌ BAD: Document all flags in SKILL.md
search-conversations supports --text, --both, --after DATE, --before DATE, --limit N

# ✅ GOOD: Reference --help
search-conversations supports multiple modes and filters. Run --help for details.

Use cross-references:

# ❌ BAD: Repeat workflow details
When searching, dispatch subagent with template...
[20 lines of repeated instructions]

# ✅ GOOD: Reference other skill
Always use subagents (50-100x context savings). REQUIRED: Use [other-skill-name] for workflow.

Compress examples:

# ❌ BAD: Verbose example (42 words)
your human partner: "How did we handle authentication errors in React Router before?"
You: I'll search past conversations for React Router authentication patterns.
[Dispatch subagent with search query: "React Router authentication error handling 401"]

# ✅ GOOD: Minimal example (20 words)
Partner: "How did we handle auth errors in React Router?"
You: Searching...
[Dispatch subagent → synthesis]

Eliminate redundancy:

• Don’t repeat what’s in cross-referenced skills
• Don’t explain what’s obvious from command
• Don’t include multiple examples of same pattern

Verification:

wc -w skills/path/SKILL.md
# getting-started workflows: aim for <150 each
# Other frequently-loaded: aim for <200 total

Name by what you DO or core insight:

• ✅ condition-based-waiting > async-test-helpers
• ✅ using-skills not skill-usage
• ✅ flatten-with-flags > data-structure-refactoring
• ✅ root-cause-tracing > debugging-techniques

Gerunds (-ing) work well for processes:

• creating-skills, testing-skills, debugging-with-logs
• Active, describes the action you’re taking

5. Cross-Referencing Other Skills

When writing documentation that references other skills:

Use skill name only, with explicit requirement markers:

• ✅ Good: **REQUIRED SUB-SKILL:** Use superpowers:test-driven-development
• ✅ Good: **REQUIRED BACKGROUND:** You MUST understand superpowers:systematic-debugging
• ❌ Bad: See skills/testing/test-driven-development (unclear if required)
• ❌ Bad: @skills/testing/test-driven-development/SKILL.md (force-loads, burns context)

Why no @ links: @ syntax force-loads files immediately, consuming 200k+ context before you need them.

Flowchart Usage

digraph when_flowchart {
    "Need to show information?" [shape=diamond];
    "Decision where I might go wrong?" [shape=diamond];
    "Use markdown" [shape=box];
    "Small inline flowchart" [shape=box];

    "Need to show information?" -> "Decision where I might go wrong?" [label="yes"];
    "Decision where I might go wrong?" -> "Small inline flowchart" [label="yes"];
    "Decision where I might go wrong?" -> "Use markdown" [label="no"];
}

Use flowcharts ONLY for:

• Non-obvious decision points
• Process loops where you might stop too early
• “When to use A vs B” decisions

Never use flowcharts for:

• Reference material → Tables, lists
• Code examples → Markdown blocks
• Linear instructions → Numbered lists
• Labels without semantic meaning (step1, helper2)

See graphviz-conventions.dot in this directory for graphviz style rules.

Visualizing for your human partner: Use render-graphs.js in this directory to render a skill’s flowcharts to SVG:

./render-graphs.js ../some-skill           # Each diagram separately
./render-graphs.js ../some-skill --combine # All diagrams in one SVG

Code Examples

One excellent example beats many mediocre ones

Choose most relevant language:

• Testing techniques → TypeScript/JavaScript
• System debugging → Shell/Python
• Data processing → Python

Good example:

• Complete and runnable
• Well-commented explaining WHY
• From real scenario
• Shows pattern clearly
• Ready to adapt (not generic template)

Don’t:

• Implement in 5+ languages
• Create fill-in-the-blank templates
• Write contrived examples

You’re good at porting - one great example is enough.

File Organization

Self-Contained Skill

defense-in-depth/
  SKILL.md    # Everything inline

When: All content fits, no heavy reference needed

Skill with Reusable Tool

condition-based-waiting/
  SKILL.md    # Overview + patterns
  example.ts  # Working helpers to adapt

When: Tool is reusable code, not just narrative

Skill with Heavy Reference

pptx/
  SKILL.md       # Overview + workflows
  pptxgenjs.md   # 600 lines API reference
  ooxml.md       # 500 lines XML structure
  scripts/       # Executable tools

When: Reference material too large for inline

The Iron Law (Same as TDD)

NO SKILL WITHOUT A FAILING TEST FIRST

This applies to NEW skills AND EDITS to existing skills.

Write skill before testing? Delete it. Start over. Edit skill without testing? Same violation.

No exceptions:

• Not for “simple additions”
• Not for “just adding a section”
• Not for “documentation updates”
• Don’t keep untested changes as “reference”
• Don’t “adapt” while running tests
• Delete means delete

REQUIRED BACKGROUND: The superpowers:test-driven-development skill explains why this matters. Same principles apply to documentation.

Testing All Skill Types

Different skill types need different test approaches:

Discipline-Enforcing Skills (rules/requirements)

Examples: TDD, verification-before-completion, designing-before-coding

Test with:

• Academic questions: Do they understand the rules?
• Pressure scenarios: Do they comply under stress?
• Multiple pressures combined: time + sunk cost + exhaustion
• Identify rationalizations and add explicit counters

Success criteria: Agent follows rule under maximum pressure

Technique Skills (how-to guides)

Examples: condition-based-waiting, root-cause-tracing, defensive-programming

Test with:

• Application scenarios: Can they apply the technique correctly?
• Variation scenarios: Do they handle edge cases?
• Missing information tests: Do instructions have gaps?

Success criteria: Agent successfully applies technique to new scenario

Pattern Skills (mental models)

Examples: reducing-complexity, information-hiding concepts

Test with:

• Recognition scenarios: Do they recognize when pattern applies?
• Application scenarios: Can they use the mental model?
• Counter-examples: Do they know when NOT to apply?

Success criteria: Agent correctly identifies when/how to apply pattern

Reference Skills (documentation/APIs)

Examples: API documentation, command references, library guides

Test with:

• Retrieval scenarios: Can they find the right information?
• Application scenarios: Can they use what they found correctly?
• Gap testing: Are common use cases covered?

Success criteria: Agent finds and correctly applies reference information

Common Rationalizations for Skipping Testing

All of these mean: Test before deploying. No exceptions.

Match the Form to the Failure

Before writing guidance, classify the baseline failure. The form that bulletproofs one failure type measurably backfires on another.

Why prohibitions backfire on shaping problems: under a competing incentive (“make the prompt self-contained”), agents negotiate with “don’t X”. In head-to-head wording tests on dispatch-prompt guidance, the prohibition arm produced clearly more of the unwanted content than the recipe arm (fully separated distributions), and trended worse than even the no-guidance control — micro-test your own case rather than assuming, but never reach for the prohibition by default. A recipe leaves nothing to negotiate: the output matches the stated shape or it doesn’t.

Rules for whichever form you pick:

• No nuance clauses. “Don’t X unless it matters” reopens the negotiation — appending a single nuance clause to a winning recipe degraded it from consistent to noisy in the same wording tests. Express a real exception as its own conditional on an observable predicate.
• Exemption clauses don’t scope. “This limit doesn’t apply to code blocks” still suppresses code blocks. If part of the output must be exempt, restructure so the rule can’t reach it.

Bulletproofing Skills Against Rationalization

Skills that enforce discipline (like TDD) need to resist rationalization. Agents are smart and will find loopholes when under pressure.

Scope: this toolkit is for discipline failures — an agent that knows the rule and skips it under pressure. For wrong-shaped output or omitted elements, prohibition-based bulletproofing backfires; use the forms in Match the Form to the Failure instead.

Psychology note: Understanding WHY persuasion techniques work helps you apply them systematically. See persuasion-principles.md for research foundation (Cialdini, 2021; Meincke et al., 2025) on authority, commitment, scarcity, social proof, and unity principles.

Close Every Loophole Explicitly

Don’t just state the rule - forbid specific workarounds:

No exceptions:

• Don’t keep it as “reference”
• Don’t “adapt” it while writing tests
• Don’t look at it
• Delete means delete

</Good>

### Address "Spirit vs Letter" Arguments

Add foundational principle early:

```markdown
**Violating the letter of the rules is violating the spirit of the rules.**

This cuts off entire class of “I’m following the spirit” rationalizations.

Build Rationalization Table

Capture rationalizations from baseline testing (see Testing section below). Every excuse agents make goes in the table:

| Excuse | Reality |
|--------|---------|
| "Too simple to test" | Simple code breaks. Test takes 30 seconds. |
| "I'll test after" | Tests passing immediately prove nothing. |
| "Tests after achieve same goals" | Tests-after = "what does this do?" Tests-first = "what should this do?" |

Create Red Flags List

Make it easy for agents to self-check when rationalizing:

## Red Flags - STOP and Start Over

- Code before test
- "I already manually tested it"
- "Tests after achieve the same purpose"
- "It's about spirit not ritual"
- "This is different because..."

**All of these mean: Delete code. Start over with TDD.**

Update SDO for Violation Symptoms

Add to description: symptoms of when you’re ABOUT to violate the rule:

description: use when implementing any feature or bugfix, before writing implementation code

RED-GREEN-REFACTOR for Skills

Follow the TDD cycle:

RED: Write Failing Test (Baseline)

Run pressure scenario with subagent WITHOUT the skill. Document exact behavior:

• What choices did they make?
• What rationalizations did they use (verbatim)?
• Which pressures triggered violations?

This is “watch the test fail” - you must see what agents naturally do before writing the skill.

GREEN: Write Minimal Skill

Write skill that addresses those specific rationalizations. Don’t add extra content for hypothetical cases.

Run same scenarios WITH skill. Agent should now comply.

REFACTOR: Close Loopholes

Agent found new rationalization? Add explicit counter. Re-test until bulletproof.

Micro-Test Wording Before Full Scenarios

Full pressure-scenario runs are the final gate, but they are slow and expensive per iteration. Verify the wording itself first with micro-tests:

1. One fresh-context sample per call — a raw API call, or a single-shot subagent if you don’t have API access. System prompt = the realistic context the guidance will live in (the full skill or prompt template, not the guidance in isolation); user message = a task that tempts the failure.
2. Always include a no-guidance control. If the control doesn’t exhibit the failure, there is nothing to fix — stop, don’t author the guidance.
3. 5+ reps per variant. Single samples lie.
4. Manually read every flagged match. Score programmatically if you like, but template echoes and quoted counter-examples masquerade as hits; automated counts alone overstate both failure and success.
5. Variance is a metric. When guidance lands, reps converge on the same shape. Five different interpretations across five reps means the wording isn’t binding — tighten the form before adding words.

Micro-tests verify wording; they do not replace pressure scenarios for discipline skills.

Testing methodology: See testing-skills-with-subagents.md (opens in a new tab) for the complete testing methodology:

• How to write pressure scenarios
• Pressure types (time, sunk cost, authority, exhaustion)
• Plugging holes systematically
• Meta-testing techniques

Anti-Patterns

❌ Narrative Example

“In session 2025-10-03, we found empty projectDir caused…” Why bad: Too specific, not reusable

❌ Multi-Language Dilution

example-js.js, example-py.py, example-go.go Why bad: Mediocre quality, maintenance burden

❌ Code in Flowcharts

step1 [label="import fs"];
step2 [label="read file"];

Why bad: Can’t copy-paste, hard to read

❌ Generic Labels

helper1, helper2, step3, pattern4 Why bad: Labels should have semantic meaning

STOP: Before Moving to Next Skill

After writing ANY skill, you MUST STOP and complete the deployment process.

Do NOT:

• Create multiple skills in batch without testing each
• Move to next skill before current one is verified
• Skip testing because “batching is more efficient”

The deployment checklist below is MANDATORY for EACH skill.

Deploying untested skills = deploying untested code. It’s a violation of quality standards.

Skill Creation Checklist (TDD Adapted)

IMPORTANT: Create a todo for EACH checklist item below.

RED Phase - Write Failing Test:

• Not done: Create pressure scenarios (3+ combined pressures for discipline skills)
• Not done: Run scenarios WITHOUT skill - document baseline behavior verbatim
• Not done: Identify patterns in rationalizations/failures

GREEN Phase - Write Minimal Skill:

• Not done: Name uses only letters, numbers, hyphens (no parentheses/special chars)
• Not done: YAML frontmatter with required name and description fields (max 1024 chars; see spec (opens in a new tab))
• Not done: Description starts with “Use when…” and includes specific triggers/symptoms
• Not done: Description written in third person
• Not done: Keywords throughout for search (errors, symptoms, tools)
• Not done: Clear overview with core principle
• Not done: Address specific baseline failures identified in RED
• Not done: Guidance form matches the failure type (see Match the Form to the Failure)
• Not done: For behavior-shaping guidance: wording micro-tested against a no-guidance control (5+ reps, every flagged match read manually) — N/A for pure reference skills
• Not done: Code inline OR link to separate file
• Not done: One excellent example (not multi-language)

REFACTOR Phase - Close Loopholes:

• Not done: Identify NEW rationalizations from testing
• Not done: Add explicit counters (if discipline skill)
• Not done: Build rationalization table from all test iterations
• Not done: Create red flags list
• Not done: Re-test until bulletproof

Quality Checks:

• Not done: Small flowchart only if decision non-obvious
• Not done: Quick reference table
• Not done: Common mistakes section
• Not done: No narrative storytelling
• Not done: Supporting files only for tools or heavy reference

Deployment:

• Not done: Commit skill to git and push to your fork (if configured)
• Not done: Consider contributing back via PR (if broadly useful)

Discovery Workflow

How future agents find your skill:

1. Encounters problem (“tests are flaky”)
2. Searches skills (greps descriptions, browses categories)
3. Finds SKILL (description matches)
4. Scans overview (is this relevant?)
5. Reads patterns (quick reference table)
6. Loads example (only when implementing)

Optimize for this flow - put searchable terms early and often.