Template for a workflow-* skill (a production/research process, not an ad format)

Directory: ~/.openclaw/workspace/skills/workflow-<name>/ → symlink into ~/.openclaw/workspace/.claude/skills/.

workflow-<name>/
├── SKILL.md                 # ≤250 lines, the 9 sections below
├── references/
│   ├── procedure.md         # step-by-step detail, prompts/templates, worked ILLUSTRATIVE example
│   └── sources.md           # davidaistar videos / diffs it came from, what we kept / changed, our receipts
├── evals/evals.json         # 4-6 cases: prompt → must / must_not
└── memory/learnings.jsonl   # seeded receipts; append after every run

Frontmatter: name: workflow-<name>, description: "<what> . Use when <triggers>. Inherits knowledge/ad-formats/_FORMAT-CORE.md and _MODEL-REGISTRY.md." (quoted).

SKILL.md sections (##, in order)

1. Read first — core, registry, references, brand files.
2. When to use / when not
3. Inputs — what must exist before starting (files, approvals, assets).
4. Outputs — exact files/dirs produced, naming.
5. Process — numbered steps; each step names its tool/script and whether it is free or paid. - First bullet: Step 0 — combo plan (core §0): plan.py --format <this skill> → paste the combo block (all combo_* keys incl. combo_ad_prefix) → tags.py check PASS.
6. Tools & models — ROLES from _MODEL-REGISTRY.md (never hard-code a model in prose without its role); what is built vs NOT BUILT (needs Fish's go).
7. Gates & QA checklist
8. Failure modes
9. Sources