Skill Writing

Skill Writing

Bootstrap a project-local skill with a small observable contract, then symlink it into .claude/skills/ so Claude Code can discover it. Keep invariant workflow guidance in SKILL.md; move deterministic mechanics and conditional detail into scripts and references.

Model Optimization

Optimize every new skill and its content for GPT-6 Astra and Claude Fable 5.1. The summaries below are reminders, not substitutes for the live guides. Read both guides before designing or writing a complex, long-running, multi-tool, or orchestration-heavy skill because their recommendations may evolve.

• GPT-6 Astra prompting guidance: Complete authorized work under stated assumptions; make user-instruction precedence over skills explicit; specify writing and delegation preferences; and keep verification proportional to the change.
• Claude Fable 5.1 prompting guidance: Calibrate effort with evals; request progress updates; batch independent tool calls; preserve decisions across compaction; verify changing facts; and finish the requested scope with targeted edits and proportionate tests.

Arguments

• skill-name (required): kebab-case name (e.g., my-skill). Stop if missing or invalid.

Reject --global, explicit destination paths, and other scope overrides. The invocation working directory is the only supported scope.

Repository Catalog Guard

Before resolving project-local paths, read the repository instructions applicable to the invocation working directory. If they define a source catalog and lifecycle for skill creation, stop this workflow and follow that repository-owned workflow. Do not create .agents/skills/ or .claude/skills/ paths in that repository.

Resolved Paths

Set <scope> to the working directory where the skill was invoked. Create the source at <scope>/.agents/skills/<name>/ and the Claude Code symlink at <scope>/.claude/skills/<name>. Do not redirect the scope to the repository root when invoked from a nested project or workspace. Never create or modify a skill under ~/.agents, ~/.claude, ~/.codex, or another global installation directory.

The symlink target is always the relative path ../../.agents/skills/<name>.

Skill Layout

<name>/
├── SKILL.md       # Required: frontmatter + lean workflow (aim for <500 lines)
├── agents/
│   └── openai.yaml # Required: Codex invocation policy derived from SKILL.md
├── scripts/       # Optional: helper code (prefer TypeScript via bun run; Python via uv)
├── references/    # Optional: long-form docs loaded on demand
└── assets/        # Optional: templates / fonts / images used in OUTPUT (never loaded into context)

Agents load skills via progressive disclosure, in three stages:

1. Discovery — only name + description are visible at startup. Front-load triggers in description.
2. Activation — the full SKILL.md body is read once a task matches.
3. Execution — scripts/ run without being read into context; references/ are read only when SKILL.md explicitly links to them.

Keep SKILL.md focused on workflow. Push bulk into scripts/ (deterministic logic) or references/ (documentation).

Authoring Contract

Before choosing a layout, separate the content into:

• Invariants that every valid execution must preserve.
• Preferred defaults that explicit user intent or repository evidence may override.
• Conditional examples and references loaded only when their branch is active.

Define the outcome, authority boundaries, stopping conditions, and completion evidence. Do not prescribe an identical execution path when several safe paths satisfy the same contract. For user-facing workflows, also define which kickoff, progress, decision, blocker, and completion events deserve a message and the smallest useful shape for each.

When a workflow hands long-winded work to background jobs or agents, make polished progress monitoring part of the contract. Judge whether the work is long-winded from its expected runtime and uncertainty, fan-out or waves, meaningful milestones, and the visibility the host already provides rather than a universal time cutoff. Name the main agent as the reporting owner and follow the long-winded background-work guidance in references/writing-great-skills.md.

Express outcomes, invariants, and completion evidence as positive, observable acceptance criteria; when affirmative evidence is available, require it instead of accepting only the absence of listed failures or adding negative examples, inverse restatements, or long blacklists. Retain a negative instruction only for an explicit user-requested exclusion or when it is the clearest concise guard for a consequential safety, authority, destructive-action, scope, or likely model failure boundary that positive criteria cannot enforce equivalently.

When to Split Content

Use scripts/ when

• The same code would be rewritten on every invocation (e.g., PDF rotate, JSON transform, curl wrapper).
• Determinism matters more than flexibility (parsing, validation, codegen, idempotent setup).
• A shell pipeline grows past ~5 lines or needs real error handling.
• A long heredoc keeps appearing inside SKILL.md.

Scripts are token-efficient: the agent invokes them without reading them. Document the CLI signature in SKILL.md and leave the implementation in scripts/.

Prefer scripts/*.ts run with bun run scripts/<name>.ts, unless there is a good reason TypeScript is the wrong fit for the helper. Python is also a good choice for data, text, and file processing; run Python helpers through uv run scripts/<name>.py, not raw python or python3.

Keep Bash helpers compatible with macOS /bin/bash 3.2.

Use references/ when

• A topic exceeds ~100 lines of prose, examples, or schemas.
• Content is conditionally relevant (variant-, framework-, or domain-specific) — splitting keeps irrelevant context out.
• Detailed API surfaces, DB schemas, policies, or large templates would otherwise dominate SKILL.md.
• A long explanation is needed only on one branch.

Rules of thumb:

• One level deep — link references/placeholder.md directly from SKILL.md, never reference-to-reference.
• Files >100 lines: include a table of contents at the top.
• Files >10k words: document grep patterns in SKILL.md so the agent can locate sections without reading the whole file.
• No duplication — each fact lives in SKILL.md or a reference, never both. Keep skills self-contained rather than sharing references across independently installed skills.
• For every reference, write one line in SKILL.md that says when to read it.

Reference organization patterns

Pattern A — High-level guide + topical references

SKILL.md
references/
├── forms.md
├── api.md
└── examples.md

SKILL.md teaches the happy path; references hold deep-dive material.

Pattern B — Domain or variant split

SKILL.md           # workflow + selection logic
references/
├── aws.md
├── gcp.md
└── azure.md

The agent reads only the variant the user picked — irrelevant providers never enter context.

Pattern C — Conditional details

Inline the basic case in SKILL.md, link advanced files for edge cases (tracked-changes.md, ooxml.md, etc.).

Exclude runtime-irrelevant files

Do not add repository-style support files (README.md, INSTALLATION.md, CHANGELOG.md, or QUICK_REFERENCE.md), authoring notes, test logs, scratch files, or anything the agent will not use at runtime.

Workflow

1. Read Format Docs

Resolve scripts/fetch-agentskills-spec.sh relative to this skill directory, run it once, and read the returned file completely. The helper reuses an integrity-valid specification for 24 hours, conditionally revalidates older entries, and may return a cache validated within seven days when live retrieval fails. Set AGENTSKILLS_CACHE_DIR when the default user cache location is unavailable or unwritable.

Use --refresh for explicitly latest or change-sensitive work, disputed portable-format guidance, or a conflict with validator behavior. A stale result is usable only after reading it; disclose its validation timestamp and retrieval failure in the completion report. If the helper cannot return a valid file, stop before writing. Never fetch the agentskills.io specification directly or create its cache in a repository or skill installation.

Fetch the current Claude Code frontmatter reference with WebFetch. Confirm field shapes, naming rules, and progressive-disclosure conventions from both sources; do not guess because the formats evolve.

2. Validate

• Reject names that are not kebab-case or collide with an existing skill at the resolved path.
• Confirm <scope> is the invocation working directory and is not a home-level or global skill directory.
• Stop if <scope>/.agents/skills/<name>/ or <scope>/.claude/skills/<name> already exists.

3. Read the Authoring Guide and Define the Contract and Layout

Read references/writing-great-skills.md before choosing the contract or layout. It defines the predictability levers and the prose-versus-code-or-schema decision.

Then define the observable outcome, invariants, preferred defaults, authority, routing, stop conditions, and completion evidence. While defining routing, identify every skill that the workflow requires, invokes, or hands off to on any supported branch. Exclude suggestions, examples, related-skill references, and underlying tool capabilities. Decide what belongs where:

• Will the workflow invoke helper code? → Prefer scripts/<name>.ts run with bun run; use scripts/<name>.py through uv run when Python is a better fit.
• Machine-consumed schema with a real validator? → Bundle both and document the validation route.
• Reference-only schema documentation, long examples, variant guides, or domain knowledge? → references/<topic>.md
• Templates or files the skill writes into the user's output? → assets/
• None of the above? → ship only the required SKILL.md and agents/openai.yaml.

Sketch the directory tree first, then create only the subdirectories the layout actually needs.

4. Create the Skill

mkdir -p "<scope>/.agents/skills/<name>/agents"
# Add only the subdirectories the layout calls for:
# mkdir -p "<scope>/.agents/skills/<name>/scripts"
# mkdir -p "<scope>/.agents/skills/<name>/references"

Write <scope>/.agents/skills/<name>/SKILL.md with:

• Frontmatter sorted alphabetically, with description last. Front-load discovery-time trigger phrases in description.

• A skill-dependencies array when routing identified dependencies. Use bare names for skills in the same repository and ORG/REPO#SKILL for external skills. Sort by the target skill name (the bare name or substring after #), then by the complete identifier. Require unique strings, resolve every bare dependency in the same repository, and exclude the owning skill as a bare dependency. External repository existence is not validated. Omit the field when no dependencies exist.

• A short # Title.

• A one-line summary of what the skill does.

• Add disable-model-invocation: true or user-invocable: false only when the skill differs from Claude's defaults. Omit disable-model-invocation: false and user-invocable: true because absence already expresses those values.

• Set coordination: exempt only when the skill's declared default workflow writes no repository files or only repository metadata. When selected, add this ordinary prose declaration to the new skill's body; the fence below is documentation for this authoring skill, not its own declaration:

  This skill is coordination-exempt: skip the ai-coord gate for its declared work.
  

  Explicitly authorized escalation beyond the declared behavior re-enters the gate.

• ## Arguments (if any) and a lean imperative workflow. Use fixed steps only when order matters; otherwise state the contract and let repository evidence guide execution.

• Explicit links to every references/ file the workflow may need, each with a one-line note describing when to read it.

• CLI signatures for any bundled scripts, including the runtime command (bun run scripts/<name>.ts or uv run scripts/<name>.py), so the agent can call them without reading them.

Use imperative prose and resolve bundled references/, scripts/, examples/, and assets/ paths relative to the owning skill directory.

Aim for SKILL.md under 500 lines. If a section grows past ~50 lines and is not core workflow, move it to references/ and link it.

Write <scope>/.agents/skills/<name>/agents/openai.yaml with:

policy:
  allow_implicit_invocation: true

Set allow_implicit_invocation to the inverse of SKILL.md disable-model-invocation. If later adding Codex UI metadata or MCP/tool dependencies, merge them into the same file and keep the policy.

5. Create the Claude Code Symlink

Always create a relative symlink so Claude Code picks the skill up from its own discovery path:

mkdir -p "<scope>/.claude/skills"
ln -s "../../.agents/skills/<name>" "<scope>/.claude/skills/<name>"

6. Verify

• Patch tooling creates files at mode 0644. Before the first verification run, chmod 755 every executable under scripts/ and tests/ (a scaffolded test failing its first run with Permission denied (os error 13) is this cause).
• test -f "<scope>/.agents/skills/<name>/SKILL.md"
• test -f "<scope>/.agents/skills/<name>/agents/openai.yaml"
• readlink "<scope>/.claude/skills/<name>" resolves to the source directory.
• test -x every scripts/* and tests/* executable so a missed chmod fails loudly instead of surfacing later as a permission error.
• ai-skillet doctor --root "<scope>/.agents/skills/<name>" exits 0. This is the canonical local schema and policy gate.
• Finish with ### 🧩 Skill created: <name>, a tree of created paths, and ### ✅ Verified with the exact checks. Link both absolute source and symlink paths.
• Offer to commit the new skill. When the host project's standing instructions require prompt commits, commit without further prompting.

Notes

• The skills CLI parses SKILL.md frontmatter as YAML before publishing. A colon followed by a space inside a plain scalar, such as leave: freeze in description, makes that parser fail. Use an em dash or another safe separator, or quote the entire scalar.
• Keep helper stdout, commands, paths, frontmatter, and generated skill content undecorated unless that skill's own output contract requires otherwise.