skills-best-practices

Opinionated best practices for building Agent Skills - SKILL.md structure, frontmatter, description writing, progressive disclosure, testing, distribution. Use when creating or reviewing a skill, or debugging why one will not trigger.

Opinionated guide to building Agent Skills for any agent - distilled from the Agent Skills open standard, Anthropic's official guidance, and production experience, with deviations from the official line marked where they occur. Skills are folders (often just a single file) containing instructions, scripts, and resources that teach an agent how to handle specific tasks.

Quick Start

A minimal skill is a directory with a SKILL.md file:

text

my-skill/
├── SKILL.md          # Required - instructions with YAML frontmatter
├── references/       # Optional - detailed docs loaded on demand
├── scripts/          # Optional - executable code
└── assets/           # Optional - templates, fonts, icons

Minimal SKILL.md:

yaml

---
name: my-skill-name
description: What it does. Use when [specific triggers].
---

# My Skill Name

[Instructions here]

Only name and description are required in frontmatter.

Core Design Principles

Single File vs. references/ (Most Important)

Default to a single SKILL.md. One file can be pasted to a person, gisted, embedded in a CLI binary, and printed by a <tool> skill subcommand - a directory cannot. Split into references/ only when both hold:

1. Conditional loading: a meaningful chunk of content is needed by only a subset of invocations (e.g. a tracked-changes doc most DOCX tasks never touch). If every invocation reads everything anyway, splitting adds Read round-trips and costs shareability while saving nothing.
2. Size pressure: the body exceeds the recommended budget below.

Distribution is a veto. If the skill must travel as one file - shipped inside a CLI, printed by a command, shared by paste - stay single-file regardless of size and condense instead. Condensing means cutting redundancy, filler, and over-explanation while preserving every load-bearing instruction; losing substance to hit a line count is the failure mode, not the fix. See the single-file CLI-embedded pattern under Patterns.

Size guidance (opinionated thresholds drawn from experience, not enforced spec limits) - measure with wc -c SKILL.md. Chars track token cost closely (~4 chars per token); line counts are not a metric - identical content varies 2x in lines by formatting style:

Tier	Chars	Beyond it
Recommended	25k	Condense carefully; split only if the conditional-loading test passes
Hard ceiling	50k	Must condense or split

Official Anthropic guidance says to split at 500 lines. That advice assumes registry-installed skills with rarely-needed subtopics, and measures size in a unit that formatting distorts - this skill deliberately deviates on both.

When a skill does split, information loads in three levels:

Level	When Loaded	Token Cost	Content
1: Metadata	Always (startup)	~100 tokens	name + description from frontmatter
2: Instructions	When skill triggers	<5k tokens (recommended)	SKILL.md body
3: Resources	As needed	Effectively unlimited	Bundled files, scripts

Reference detail files from SKILL.md so they load only when the task requires them:

markdown

## Advanced features
- **Form filling**: See [FORMS.md](FORMS.md)
- **API reference**: See [reference.md](reference.md)

Composability

Skills work alongside other skills. Don't assume yours is the only one loaded.

Portability

Skills work across Claude.ai, Claude Code, API, and Agent SDK without modification (if dependencies are available).

Writing the Description (Critical)

The description is the single most important field - it determines when your skill activates. Claude uses it to decide relevance from potentially 100+ available skills.

Rules

• Write in third person ("Processes files..." - first or second person breaks discovery)
• Include WHAT it does + WHEN to use it
• Max 1024 characters, no XML angle brackets
• Be slightly "pushy" - Claude tends to undertrigger rather than overtrigger
• Include specific trigger phrases users would naturally say, plus file types where relevant
• Write natural prose, not keyword dumps - matching is semantic, so a long "Triggers on X, Y, Z..." list adds little over a clear sentence
• If the skill depends on an MCP server, name it ("...via MCP. Requires Linear MCP server connected.")

Good vs Bad

yaml

# GOOD - specific, actionable, includes triggers
description: Extract text and tables from PDF files, fill forms, merge
  documents. Use when working with PDF files or when the user mentions
  PDFs, forms, or document extraction.

# BAD - too vague
description: Helps with documents.

# BAD - missing triggers
description: Creates sophisticated multi-page documentation systems.

Negative Triggers

When a skill overtriggers, add boundaries directly in the description:

yaml

description: Advanced data analysis for CSV files. Use for statistical
  modeling, regression, clustering. Do NOT use for simple data
  exploration (use data-viz skill instead).

Manually-Invoked Skills

A skill with disable-model-invocation: true never auto-triggers - its description shows only in the / menu, so trigger phrases do nothing for it. Write a plain one-line summary and skip the trigger-tuning.

Frontmatter Reference

Required Fields

Field	Rules
name	Kebab-case, max 64 chars, lowercase + numbers + hyphens only. No "claude" or "anthropic"
description	Non-empty, max 1024 chars, no XML tags. WHAT + WHEN

The agentskills.io standard and the Claude API require both fields. Claude Code is more lenient: name falls back to the directory name, and description falls back to the first markdown paragraph. Write both anyway for portability.

The spec also defines optional license, compatibility, and metadata fields. compatibility is capped at 500 characters and states environment requirements (intended product, system packages, network access).

Optional Fields (Claude Code)

Field	Purpose
argument-hint	Autocomplete hint, e.g. [issue-number]
when_to_use	Extra trigger context, appended to description in the skill listing
arguments	Named positional arguments for $name substitution (space-separated string or list)
disable-model-invocation	true = only user can invoke (for deploy, commit)
user-invocable	false = hidden from / menu (background knowledge)
allowed-tools	Pre-approves tools (no permission prompt) for the current turn; space-separated, e.g. Read Grep Glob. In the spec allowlist but tagged (Experimental)
disallowed-tools	Removes tools from Claude's pool while the skill is active; clears on your next message
model	Override model for this skill; accepts inherit. Lasts the current turn only
effort	Override effort level: low, medium, high, xhigh, max
context	fork = run in isolated subagent
agent	Subagent type when context: fork (e.g. Explore, Plan)
background	false opts a forked skill out of background execution (v2.1.218+)
shell	bash (default) or powershell
hooks	Hooks scoped to this skill's lifecycle
paths	Glob patterns limiting when skill activates

Publishing caveat: every field above except allowed-tools is Claude Code-specific. They work in Claude Code at runtime, but the official agentskills validate spec validator rejects them - it allows only name, description, license, compatibility, metadata, allowed-tools, with no relax flag. If your repo or CI runs that validator (most ClawHub-publishing repos do), a skill using these fields fails validation unless you strip them from the copy you validate/publish. The ClawHub registry itself tends to tolerate extra top-level fields on publish, but the reference validator in your pipeline will not. See Validate Against the Spec.

Naming Conventions

The name (and its folder) must: be 1-64 chars; use only lowercase letters, numbers, and hyphens; not start or end with a hyphen; not contain consecutive hyphens (--); and match the parent directory name. Anthropic surfaces also reject the reserved words claude and anthropic.

Prefer gerund form for clarity:

• processing-pdfs, analyzing-spreadsheets, managing-databases
• Also acceptable: pdf-processing, process-pdfs
• Avoid: helper, utils, tools, documents

Claude Code Specifics

Official docs cover most Claude Code skill behavior: the skills docs (invocation control, argument substitution, discovery and priority, tool permissions, skillOverrides, context budget), the commands reference for the current bundled-skills roster (it churns every few releases - never hardcode it), and the settings reference. Below is only what those docs miss or what bites in practice.

Dynamic-Injection Footgun

Claude Code preprocesses SKILL.md at load: an exclamation mark immediately touching a backticked command executes that command before Claude sees the content (dynamic context injection). The preprocessor is not markdown-aware:

• A literal example executes at load even inside a code fence or inline code span, and a failing placeholder command errors the whole skill at load
• The inline form fires only at line start or after whitespace; a prefix defuses it (KEY= before the ! leaves it literal)
• A fence opened with ! right after the backticks is the multi-line form and is equally live
• references/ files are read with the Read tool and never preprocessed - the only safe home for live examples. In a SKILL.md, break the !-to-backtick adjacency instead (wrap the ! in its own code span, as this section does)
• "disableSkillShellExecution": true in settings disables execution for user/project/plugin skills

Undocumented Behavior

• display-name, default-enabled, and fallback frontmatter keys exist but are absent from the official frontmatter table
• Frontmatter keys parse case-insensitively - kebab-case, snake_case, and camelCase resolve to the same field; boolean fields also accept yes/no/on/off/1/0 (v2.1.218+)

Behavior That Bites

• context: fork skills run in the background by default since v2.1.218 (background: false opts out); backgrounded forks get a narrower tool set and their edits bypass checkpoints, so /rewind cannot undo them. Explore/Plan forks skip CLAUDE.md and git status; since v2.1.198 Explore inherits the session model
• Skills stack: /skill-a /skill-b args in one message loads up to six skills (v2.1.199+)
• permissions.additionalDirectories does not load skills from those directories - only the --add-dir flag and /add-dir command do
• The allowed-tools grant lasts the current turn - it clears when the user sends their next message, not when the skill "finishes"
• The /command name comes from the skill's directory; frontmatter name is only a display label (plugin skills excepted). Nested skills are invocable by qualified name, e.g. /apps/web:deploy
• Invoked skill content stays in context all session and is not re-read - write standing instructions, not one-time steps. After auto-compaction, each skill's most recent invocation is re-attached with its first 5,000 tokens from a shared 25,000-token budget filled most-recent-first; re-invoke to restore full content
• Skill descriptions load at startup within a listing budget of 1% of the context window; least-used descriptions drop first (names always kept), each entry capped at 1,536 chars. Diagnose with /doctor; tune via skillListingBudgetFraction, skillListingMaxDescChars, or SLASH_COMMAND_TOOL_CHAR_BUDGET

Structuring Instructions

Be Concise

Claude is smart. Only add context it doesn't already have:

markdown

# GOOD (~50 tokens)
## Extract PDF text
Use pdfplumber for text extraction:
```python
import pdfplumber
with pdfplumber.open("file.pdf") as pdf:
    text = pdf.pages[0].extract_text()

BAD (~150 tokens)

Extract PDF text

PDF files are a common file format containing text and images. To extract text, you need a library. There are many available...

text

### Avoid Too Many Options

Don't present multiple approaches unless necessary. Give one default with an escape hatch:

```markdown
# BAD: "Use pypdf, or pdfplumber, or PyMuPDF, or pdf2image..."
# GOOD: "Use pdfplumber for text extraction. For scanned PDFs needing
#        OCR, use pdf2image with pytesseract instead."

Set Degrees of Freedom

• High freedom (text guidelines): Multiple approaches valid, context-dependent
• Medium freedom (pseudocode/templates): Preferred pattern exists, some variation OK
• Low freedom (exact scripts): Operations are fragile, consistency critical

Recommended SKILL.md Structure

markdown

# Skill Name

## Quick start
[Minimal working example]

## Workflow Decision Tree
[Route to the right approach based on task type]

## Detailed Instructions
[Step-by-step for each workflow]

## Examples
[Concrete input/output pairs]

## Troubleshooting
[Common errors and fixes]

Reference Files

Keep references one level deep from SKILL.md. "Depth" means the reference chain (a file linking to a file linking to a file), not filesystem nesting - a references/ subdirectory is fine. In a chain, Claude may preview files with partial reads (head) and miss content.

markdown

# BAD: Too deep
SKILL.md -> advanced.md -> details.md -> actual info

# GOOD: One level
SKILL.md -> advanced.md (contains the info directly)
SKILL.md -> reference.md (contains the info directly)

For reference files >100 lines, include a table of contents at the top. Watch file size too: a single reference of many hundreds of lines defeats progressive disclosure even at one level deep, because Claude loads the whole file for any subtopic. Split large references by subtopic so each task pulls only what it needs.

Patterns

Sequential Workflow

markdown

## Step 1: Analyze input
Run: `python scripts/analyze.py input.pdf`

## Step 2: Validate
Run: `python scripts/validate.py fields.json`
Fix any errors before continuing.

## Step 3: Execute
Run: `python scripts/process.py input.pdf fields.json output.pdf`

Conditional Workflow (Decision Tree)

markdown

## Workflow Decision Tree
**Creating new content?** -> Follow "Creation workflow"
**Editing existing content?** -> Follow "Editing workflow"
**Reviewing content?** -> Follow "Review workflow"

Feedback Loop

markdown

1. Make edits
2. Validate: `python scripts/validate.py`
3. If validation fails -> fix issues -> go to step 2
4. Only proceed when validation passes

Checklist Pattern (for complex tasks)

markdown

Copy this checklist and track progress:
- [ ] Step 1: Analyze input
- [ ] Step 2: Create plan
- [ ] Step 3: Validate plan
- [ ] Step 4: Execute
- [ ] Step 5: Verify output

Single-File Skill Embedded in a CLI

For skills documenting a CLI tool: keep SKILL.md as one file next to the CLI source, compile it into the binary (go:embed, Rust include_str!, or equivalent), and add a <tool> skill subcommand that prints it. The printed guide always matches the installed version, and one command fetches the whole doc - playwright-cli, browser-use (browser-use skill show), and agent-browser (agent-browser skills get core) all converge on this shape. Never split such a skill into references/; condense carefully instead.

Working with MCP and Subagents

MCP provides tool access; skills provide the workflow knowledge for using those tools well. Reference MCP tools by qualified name (BigQuery:bigquery_schema, GitHub:create_issue). Skills are portable expertise; subagents are isolated execution - in Claude Code, context: fork frontmatter runs a skill inside a subagent.

Developing Skills with Claude (A/B Loop)

Build skills with two Claude instances: Claude A helps design and refine (it knows the format and what agents need); Claude B is a fresh instance with the skill loaded, tested on real tasks. Notice what context you repeatedly supply during normal work, have A capture it as a skill, test with B, bring B's specific failures back to A ("it forgot to filter test accounts"), and repeat. Iterate on observed behavior, not assumptions. For output-style skills, input/output example pairs communicate the desired style better than any description.

Scripts

When your skill includes executable code:

• Solve, don't punt: Handle errors explicitly instead of letting them fail
• Justify constants: No magic numbers - document why each value was chosen
• Prefer execution over loading: Scripts run without entering context; only output consumes tokens
• Clarify intent: "Run analyze.py" (execute) vs "See analyze.py" (read as reference)
• List dependencies in SKILL.md and verify availability

Testing

Build Evaluations First

Create evaluations before writing extensive instructions - this proves the skill solves a real problem. Run Claude on representative tasks without the skill and document the failures; build ~3 scenarios that test those gaps; measure a baseline; then write the minimum instructions needed to pass. Iterate against the baseline.

Triggering Tests

text

Should trigger:
- "Help me set up a new project in [Service]"
- "I need to create a project" (paraphrased)

Should NOT trigger:
- "What's the weather?" (unrelated)
- "Write Python code" (too generic)

Functional Tests

Test normal operations, edge cases, and out-of-scope requests. Run the same request 3-5 times to check consistency.

Debug Triggering

Ask Claude: "When would you use the [skill-name] skill?" - it quotes the description back. Adjust based on what's missing.

Validate Against the Spec

Run the official Agent Skills validator before publishing:

bash

uvx --from skills-ref agentskills validate path/to/skill

Exit 0 means valid. It checks SKILL.md format and enforces the spec's strict frontmatter allowlist (name, description, license, compatibility, metadata, allowed-tools). Most registries (e.g. ClawHub) and CI gates run this, so validating locally catches failures early. If you rely on Claude Code-only frontmatter (see the publishing caveat under Frontmatter Reference), strip those fields from the copy you validate.

Pre-Publish Checklist

Calibrate to scope: for a project-local or single-user skill, skip the triggering-accuracy and distribution-hygiene items.

• Folder and name kebab-case and matching; file is exactly SKILL.md
• Description: third person, WHAT + WHEN, specific triggers, under 1024 chars, no angle brackets
• Single file unless conditionally-loaded content justifies references/; within size budget (wc -c)
• Critical instructions at the top; working examples, not pseudocode; consistent terminology
• If split: references linked from SKILL.md, one level deep, TOC for files over 100 lines
• Scripts: explicit error handling, no unexplained constants, dependencies listed, execute-vs-read intent clear
• Triggering tested: fires on direct and paraphrased requests, silent on unrelated and similar-but-distinct ones
• Functional: normal and edge cases pass, output consistent across 3-5 runs, tested on more than one model
• No time-sensitive info, Windows-style paths, or deprecated APIs
• Spec validator exits 0 (command above)
• After upload: monitor under/over-triggering in real conversations, iterate the description, bump version on every change

Troubleshooting

Symptom	Cause	Fix
Skill never loads	Description too vague	Add specific triggers and key terms
Skill loads for wrong tasks	Description too broad	Add negative triggers, be more specific
Instructions not followed	Too verbose or buried	Put critical instructions at top, use headers
Slow/degraded responses	SKILL.md too large	Condense first; split to references/ only if content is conditionally loaded (see Single File vs. references/)
"Could not find SKILL.md"	Wrong filename	Must be exactly SKILL.md (case-sensitive)
"Invalid skill name"	Spaces or capitals	Use kebab-case: my-skill-name
Whole skill silently skipped at load	Description exceeds 1024 chars	Trim it - the loader rejects the file, not just the description
Frontmatter fails to parse	Triggers: (colon-space) or straight "quotes" inside an unquoted description value	Quote the whole value or remove the colon/quotes
A doc example runs a shell command	A ! directly touching a backticked command executes on load, even inside a code fence	Move the example to references/ or break the !-backtick adjacency (see Dynamic-Injection Footgun)

Distribution

Surface	How to Deploy
Claude.ai	Settings > Features > Upload zip
Claude Code (personal)	~/.claude/skills/<name>/SKILL.md
Claude Code (project)	.claude/skills/<name>/SKILL.md
Claude Code (plugin)	<plugin>/skills/<name>/SKILL.md
API	Upload via the Skill Management API, use via the Messages API
Enterprise	Managed settings (org-wide)

Skills don't sync across surfaces - deploy separately to each.

Using Skills with the API

Custom skills are uploaded through the Skill Management API; anthropic-type skills are pre-built by Anthropic. Both are used identically - pass them in the Messages API container parameter, each as {type, skill_id, version} where type is anthropic or custom. Up to 8 Skills per request, 30 MB max upload (all files combined), and all files must share a common root directory. Requires the code execution tool and the beta headers code-execution-2025-08-25 and skills-2025-10-02 (plus files-api-2025-04-14 for file upload/download).

Network access differs by surface. The API code execution environment has no network access and no runtime package installation - bundle dependencies or use pre-installed packages. On claude.ai, by contrast, Skills can install packages from npm and PyPI and pull from GitHub.

Also: a pause_turn stop reason signals a long-running Skill operation; reuse containers across turns via container.id; generated files come back via the Files API; changing the Skills list breaks prompt caching; Skills are not ZDR-eligible.

Security

• Only use skills from trusted sources
• No XML angle brackets in frontmatter (injection risk)
• Audit all bundled scripts and resources before using third-party skills
• Be cautious of skills that fetch from external URLs
• Documenting the dynamic-injection syntax is itself a hazard - the loader executes examples at load, even inside code fences. See the Dynamic-Injection Footgun before writing any

Additional References

• ClawHub publishing - source-mined moderation quirks: reason codes and fixes, LLM-review survival tactics, constraints absent from ClawHub's docs

Official Resources

• Agent Skills Spec
• Claude Code Skills Docs
• API Skills Guide
• Best Practices
• Anthropic Skills Repo
• Engineering Blog: Agent Skills
• Complete Guide PDF