Required OpenClaw skills

Four core skills plus one optional admin skill create a clean separation of responsibilities.

Skill	When used	What it does	Tools / files it owns
pokemon-card-centering-intake	User sends card photos or asks PulseLine to check centering	Validates request, collects front/back images, asks for missing side if needed, checks consent/retention language, starts analysis job.	Attachment handling, job manifest, intake copy, retake instructions.
pokemon-card-centering-vision	An analysis job needs measurement	Calls worker scripts for quality scoring, outer card detection, perspective correction, layout classification, inner-boundary detection, ratio calculation, and overlay rendering.	worker/*.py, artifact directories, debug output, eval harness.
pokemon-card-centering-rules	Measurements need grader interpretation	Loads PSA/CGC/BGS/SJC profiles, applies axis-specific threshold logic, returns pass/borderline/fail with uncertainty-aware explanations.	rules/*.json, rule validation tests, source notes.
pokemon-card-centering-report	Result needs to be shown to a human	Creates PulseLine short response, TCGMarketPulse detailed report sections, submission-risk language, and overlay captions.	report templates, explanation policy, JSON-to-copy mapping.
pokemon-card-centering-admin	Michael/admin reviews model performance or updates profiles	Runs golden-set evaluation, summarizes false pass/fail rates, records returned grades, and prepares rule/model update notes.	eval/*.py, feedback records, rule snapshots.

Skill file responsibilities

Each skill should be small and opinionated.

skills/pokemon-card-centering-intake/SKILL.md
skills/pokemon-card-centering-vision/SKILL.md
skills/pokemon-card-centering-rules/SKILL.md
skills/pokemon-card-centering-report/SKILL.md
skills/pokemon-card-centering-admin/SKILL.md

The skills should describe workflow, acceptance criteria, failure handling, and exact command wrappers. They should keep credentials and large model logic outside the skill file.

Skill interaction pattern

Composable rather than one mega-skill.

Intake skill decides whether the request is analyzable.
Vision skill produces measurements and overlays.
Rules skill converts measurements into grader eligibility.
Report skill formats the response for PulseLine or TCGMarketPulse.
Admin skill evaluates and improves the system over time.

Example skill behavior contracts

Useful boundaries to bake into SKILL.md files.

Intake never guesses: If the photo is cropped, blurry, angled, sleeved, or missing corners, ask for a retake before invoking expensive analysis.

Vision returns structured output: Every worker call returns JSON with confidence, warnings, measurement values, overlay paths, and debug paths.

Rules are versioned: Each grader decision includes rules_version and profile source note. Borderline measurements stay borderline.

Report language is guarded: Every response says “centering alone” and avoids final-grade guarantees.