| name | skill-creator-plus-for-openclaw |
| description | Create, improve, evaluate OpenClaw-compatible skills with structured iteration loops, eval design, review artifacts, and packaging guidance. Use when building a new skill, improving an existing skill, designing evals, comparing skill revisions, or tightening trigger descriptions and release readiness. |
Skill Creator Plus for OpenClaw
Build skills as products, not just folders. Use a loop of draft -> test -> review -> improve -> package, but keep the workflow compatible with OpenClaw instead of assuming Anthropic-only runtime features.
What this skill is for
Use this skill to do one or more of the following:
- create a new OpenClaw skill from scratch
- improve or refactor an existing skill
- design eval prompts and iteration loops for a skill
- compare two versions of a skill qualitatively or structurally
- improve a skill description so triggering is clearer
- package a finished skill into a distributable artifact
This skill is intentionally modeled after the stronger parts of Anthropic Skill Creator 2.0, but it should not assume Claude-only harnesses like claude -p, .claude/commands, or Anthropic-specific trigger-eval internals unless the user explicitly asks for that environment.
When execution automation is needed, prefer explicit OpenClaw-compatible command templates and persisted run artifacts over hidden runtime assumptions.
Core operating principle
Separate skill work into five stages:
- Frame: understand what the skill should do and when it should trigger
- Draft: create or improve the skill structure and bundled resources
- Evaluate: define realistic test prompts and compare outputs across iterations
- Improve: revise based on observed failure modes, not just vague preference
- Package: validate, package, and summarize next-step recommendations
Do not jump straight from “create a skill” to “done”. Prefer an explicit iteration loop when the skill matters.
OpenClaw compatibility rule
When borrowing ideas from Anthropic Skill Creator 2.0:
- keep the methodology
- adapt the mechanics
- do not blindly copy Claude-specific runtime assumptions
Examples:
- OK to reuse the idea of eval sets, baselines, benchmark summaries, reviewer loops, and description optimization
- Not OK to assume OpenClaw can directly use
.claude/commands, claude -p, or the exact Anthropic trigger-eval implementation
If the user wants direct compatibility with Anthropic tooling, say so explicitly and isolate that path as a separate workflow.
Workflow
1. Capture intent
Before editing files, determine:
- what the skill should enable
- what phrases or contexts should trigger it
- what outputs matter
- whether the task is subjective or objectively testable
- whether this needs a minimal draft or a full eval-driven loop
If the user already has a draft skill, inspect it before proposing changes.
2. Choose the work mode
Use one of these modes:
- draft-only: create a practical v0 quickly
- structured-build: create skill + references/scripts + test prompts
- eval-first: focus on test cases, success criteria, and benchmark design
- refactor-and-upgrade: improve an existing skill without losing its purpose
- package-and-polish: validate, package, and tighten description/structure
If the user does not specify a mode, infer it from the stage of the work.
3. Draft the skill with progressive disclosure
Keep the main SKILL.md lean and procedural.
Move heavy details into references/.
Add scripts/ only when deterministic repeated work is clearly useful.
Prefer this split:
SKILL.md: workflow, triggers, routing, constraints
references/: templates, rubrics, schemas, examples, evaluation guidance
scripts/: repeatable helpers for validation, report generation, or packaging support
4. Design realistic evals
When a skill is important enough to test, create realistic prompts that a real user would plausibly say.
Good evals should:
- test trigger boundaries
- test representative usage patterns
- test likely failure modes
- include near-miss or ambiguity cases where helpful
Avoid toy evals unless the skill itself is toy-sized.
5. Compare iterations explicitly
When improving a skill, do not only rely on intuition. Compare:
- previous vs current structure
- previous vs current description clarity
- expected trigger scenarios
- output quality on representative prompts
- whether the skill overfits to narrow examples
If exact quantitative benchmarking is not available, do a structured qualitative comparison.
6. Improve the description deliberately
The description field is a trigger surface, not just documentation.
A good description should include:
- what the skill does
- when to use it
- specific trigger contexts
- adjacent phrases users may say without naming the skill directly
Avoid descriptions that are so narrow they undertrigger, or so broad they hijack unrelated tasks.
7. Package only after validation
Before packaging:
- validate frontmatter and structure
- remove dead placeholders
- check references are actually referenced when needed
- ensure scripts/resources are justified
Then package and summarize:
- what the skill now does
- what remains weak
- what should be tested next
Recommended resource set
Read these references when needed:
references/openclaw-vs-anthropic.md for design adaptation principles
references/eval-loop.md for an OpenClaw-compatible eval workflow
references/eval-schemas.md for Anthropic-inspired eval data structures adapted to OpenClaw
references/eval-rubric.md for scoring and analyzer-style review dimensions
references/iteration-review-template.md for per-iteration review output
references/openclaw-eval-adaptation-notes.md for what to preserve vs adapt from Anthropic 2.0
references/description-optimization.md for improving trigger descriptions
references/release-checklist.md for final validation and packaging criteria
Use helper scripts only when they provide real leverage.
Current helpers:
scripts/generate_eval_skeleton.py: generate a starter evals.json skeleton for a chosen mode
scripts/summarize_skill_diff.py: compare two skill directories and flag changes likely to affect trigger/workflow/resource behavior
scripts/generate_iteration_review.py: aggregate one or more grading.json files into a starter iteration-review.json
scripts/aggregate_benchmark_summary.py: aggregate grading/review artifacts into a lightweight benchmark-summary.json
scripts/generate_grading_stub.py: create per-eval grading stub files from an evals.json
scripts/run_openclaw_skill_eval.py: prepare a structured iteration workspace with per-eval folders, run records, and grading stubs
scripts/run_skill_eval_review_cycle.py: orchestrate the prepare/finalize parts of an eval-review cycle
scripts/run_openclaw_eval_executor.py: execute prepared eval directories via an explicit command template and persist transcripts/stdout/stderr
scripts/grade_openclaw_eval.py: grade one eval directory into a heuristic but schema-compliant grading.json
scripts/improve_skill_description.py: analyze should-trigger / should-not-trigger cases and propose a stronger description rewrite
scripts/apply_description_rewrite.py: apply a chosen candidate description back into SKILL.md
Anti-patterns
Avoid these failure modes:
- creating a skill that is just a vague essay
- overfitting a skill to one conversation
- copying Anthropic runtime assumptions into OpenClaw without adaptation
- adding scripts because they feel impressive rather than necessary
- packaging a skill before checking whether it is actually coherent
- optimizing description keywords without thinking about real trigger intent
Output expectations
When helping with a skill, prefer outputs like:
- a proposed skill architecture
- a diff plan for improvements
- a realistic eval set
- a benchmark / review plan
- a packaging readiness checklist
If the user asks to create the skill, produce a usable draft first, then propose the next iteration loop rather than trying to finish the whole maturity ladder in one shot.