gupsammy/claudest/create-claw-skill

OpenClaw Skill & Command Generator

Generate well-structured OpenClaw skills or slash commands. Both are SKILL.md files with YAML frontmatter — they share the same structure but differ in how they're triggered and described. OpenClaw uses the AgentSkills spec (pi-coding-agent) with its own frontmatter fields, tool names, and path conventions distinct from Claude Code.

Phase 0: Fetch Current Documentation

Before generating, retrieve the latest OpenClaw skill documentation:

clawdocs get "tools/skills" --no-header -q

Capture any frontmatter fields or options not already listed in {baseDir}/references/frontmatter-options.md. If clawdocs is unavailable, proceed with current references — they are sufficient. If new fields appear, use them and note the update.

Phase 1: Understand Requirements

Parse $ARGUMENTS for type hint. Users are often unclear on OpenClaw-specific conventions. Interview to gather:

1. Primary objective — What should this skill do?
2. Trigger scenarios — When should it activate? What exact phrases would a user say?
3. Inputs/outputs — What does it receive and produce?
4. Complexity — Simple, standard, or complex workflow?
5. Gating needs — Does it require specific binaries, env vars, or config keys? (drives metadata.openclaw.requires.*)
6. Execution needs — Sub-agent delegation via sessions_spawn? Command dispatch (bypass model)?

Proceed to Phase 2 when at minimum Objective and Trigger Scenarios are established.

Port Mode

When the user provides an existing Claude Code skill to port (file path or pasted content), skip the interview and apply this translation sequence:

1. Read the source skill and all its supporting files (scripts, references, examples)
2. Frontmatter — Remove invalid fields (model, context, agent, allowed-tools, hooks, license). Add metadata with requires if scripts need specific binaries. See {baseDir}/references/frontmatter-options.md for valid fields.
3. Tool names — Apply the translation table from {baseDir}/references/claw-patterns.md: Bash→exec, Read→read, Write→write, Edit→edit, Glob/Grep→exec+find/rg, WebSearch→web_search, WebFetch→web_fetch, Task→sessions_spawn, AskUserQuestion→conversational asking
4. Paths — Replace $CLAUDE_PLUGIN_ROOT with {baseDir}. Remove @file injection and bang-backtick references.
5. Scripts/references — Copy as-is if they are platform-neutral Python with no Claude Code SDK dependencies. Update any internal tool name references.
6. Proceed directly to Phase 2 Step 7 (Validate) with translated content, then Phase 3 (Deliver).

Phase 2: Generate

Apply throughout generation: use imperative voice and terse phrasing because every token in a generated skill body costs budget on every invocation. Prefer instruction over example — state the rule with its reasoning so it generalizes to every input.

Initialize directory first (when creating a new skill directory, not editing an existing file):

python3 {baseDir}/scripts/init_claw_skill.py <name> --path <dir> [--resources scripts,references,assets]

Exit 0 = directory scaffolded, proceed to Step 1. Exit 1 = naming collision; ask user whether to overwrite or rename.

Step 1 — Choose type

• Skills: Trigger-rich, third-person description ("This skill should be used when..."); auto-triggered by routing
• Commands: Concise, verb-first description, under 60 chars; user-invoked via / menu
• Dispatch commands: command-dispatch: tool with command-tool — bypasses model entirely, routes directly to a named tool (rare; for pure pass-through cases)

Step 2 — Write frontmatter

Read {baseDir}/references/frontmatter-options.md for the full OpenClaw field catalog, description patterns, and the metadata single-line JSON constraint.

Key constraint: `metadata` must be a single-line JSON object on one line. Multi-line YAML mappings under metadata are not valid in OpenClaw.

Description density rules: Keep descriptions under ~400 characters / ~100 tokens (600 chars / 150 tokens absolute max) — they load every session. Per the OpenClaw cost formula, each skill costs 195 + 97 + field lengths characters in the system prompt; a 10-skill install with verbose descriptions burns significant context on routing metadata alone. Derive trigger phrases from the user's actual words in Phase 1, not paraphrases. See the token budget and trigger derivation principles in {baseDir}/references/frontmatter-options.md.

Intensional over extensional — state the rule with its reasoning rather than listing examples that imply the rule. An intensional rule generalizes to every input the skill will encounter; an extensional list only covers the shapes shown.

Step 3 — Validate description discoverability

Before writing the body, verify the description will route correctly. Mentally generate:

1. 3 should-trigger prompts — realistic user messages that should activate this skill. Include at least one naive phrasing from a user who has never heard of the skill.
2. 3 should-NOT-trigger prompts — messages in adjacent domains that are close but should not activate. These test whether the description is too broad.

Evaluate: does the description cover all should-trigger prompts? Would it plausibly reject the should-NOT-trigger prompts? If coverage is weak, revise the description — add missing trigger phrases, tighten language to exclude adjacent domains, or add a negative trigger ("Not for X").

This step catches routing misses before the rest of the skill is built. Proceed when description coverage is adequate.

Step 4 — Write body

Construction rules:

• State objective explicitly in first sentence
• Use imperative voice ("Analyze", "Generate", "Identify") — no first-person ("I will", "I am")
• Context only when necessary for understanding
• XML tags only for complex structured data
• No "When to Use This Skill" section — body loads only after triggering; routing guidance there is never read by the routing decision
• Avoid headers deeper than H3 — deep nesting signals content that belongs in references/, not SKILL.md
• {baseDir} is the path variable for skill-relative file references (substituted before model sees the skill body)

Both skills and commands follow the same body pattern:

# Name
 
Brief overview (1-2 sentences).
 
## Process
1.Step one (imperative voice)
2.Step two
3.Step three

Dynamic Content:

Note: @file injection and bang-backtick command expansion are Claude Code features specific to Claude Code's skill loader implementation — the pi-coding-agent skill loader only supports {baseDir} path substitution and does not implement these extensions. Do not use them in generated OpenClaw skills.

Step 5 — Script opportunity scan

Read {baseDir}/references/script-patterns.md and apply the five signal patterns to every workflow step in the skill being generated:

For each identified script candidate:

1. Choose the archetype from {baseDir}/references/script-patterns.md (init/validate/transform/package/query)
2. Scaffold the script in scripts/ using the Python template from {baseDir}/references/script-patterns.md
3. Wire it into SKILL.md with: trigger condition, exact invocation using exec tool, output interpretation

Wiring rule: A script reference must state when to invoke (trigger condition), how to invoke (exact command with flags), and what to do with the result (exit code handling, which output fields matter).

Scripts are invoked via the exec tool (not Bash). Reference paths using {baseDir}/scripts/script.py.

Step 6 — Check delegation

Read {baseDir}/references/claw-patterns.md for delegation patterns, sessions_spawn usage, cross-skill reference conventions, and tool group translations (Claude Code → OpenClaw tool name mapping).

Scan for existing resources before finalizing:

Review available OpenClaw skills (check ~/.openclaw/skills/ and workspace/skills/)
For each workflow step, ask: "Do we already have this?"

Common delegation patterns:

• To invoke another OpenClaw skill: tell the model to read {baseDir}/../<other-skill>/SKILL.md via the read tool, or instruct the user to type /<other-skill-name>
• For background delegation: use sessions_spawn (non-blocking; result announced back to chat)
• Documentation lookups: exec: clawdocs get "<slug>" --no-header -q

There is no Skill tool in OpenClaw — skills are invoked by the routing model, not programmatically from within another skill.

Step 7 — Validate

When generating a new skill directory (not editing an existing single file):

python3 {baseDir}/scripts/validate_claw_skill.py <skill-directory> --output json

Exit 0 = proceed to Phase 3. Exit 1 = parse the errors array; each entry has field, message, severity. Resolve all critical and major items before writing to disk.

Explain Your Choices

When presenting the generated skill/command to the user, briefly explain:

• What you set and why — "Added metadata.openclaw.requires.bins: [jq] because the skill calls jq in a subprocess"
• What you excluded and why — "Left user-invocable at default (true), command-dispatch omitted (skill routes through model)"
• Add more trigger phrases if routing misses expected inputs

Phase 3: Deliver

Output Paths

Skills are session-snapshotted — changes take effect on the next new session, not the current one.

Write and Confirm

Before writing:

Writing to: [path]
This will [create new / overwrite existing] file.
Proceed?

After Creation

Summarize what was created:

• Name and type
• Path and when it takes effect
• How to invoke/trigger
• Suggested test scenario

Publish to ClawHub

Invoke only when the user explicitly requests distribution:

clawhub publish <skill-directory> --slug <slug> --version X.Y.Z --tags latest

Exit 0 = published. Exit 1 = validation or auth failure; read stdout for details.

Phase 4: Evaluate

Score the generated skill/command:

Target: 9.0/10.0. If below, refine once addressing the weakest dimension, then deliver.