Skills

Skills are on-demand procedures the agent reads before running a workflow the first time in a session. They are markdown files — readable by any agent, no special format. Two are shipped:

Skill	Reads before	Covers
ingest	wf ingest	anti-loop sha256 check, CLI batch modes (--changed/--pending/--reclaim), claim extraction, judged effect verification (wf verify-effects, step 4b), change-set flow, recovery paths
promote	wf mine promotions + wf promote	deterministic clustering + judged refinement, dossiers, the 7-point review checklist, the compiler gate, tombstone/suppression buffer, wf utility write-back, SLOW-REGION revision proposals

Run wf skill --list to list them; wf skill <name> to print the full procedure. The procedure is the contract — it carries the anti-loop rules and the step order that the CLI only partially enforces.

$ wf skill --list
Available skills (wf skill <name> prints the full procedure):

  ingest     Ingest a raw source into the evidence fabric via `wf ingest` —
             creates source record, faithful summary, extracts claims, opens
             a change-set for review...
  promote    Run the promotion pipeline via `wf mine promotions` + `wf promote`
             — cluster experience-events deterministically, generate dossiers...

$ wf skill ingest      # prints the procedure (agents: read BEFORE the first
                       # ingest this session — it is the step contract)

The harness loop in one pass — what an agent session looks like when the fabric is installed (the always-on block arrives at session start; the agent runs these in order):

SESSION START (from the always-on block, 0 tokens)
  → wf context --task "<task>" --write-receipt   # scoped manifest + receipt id
  → [optional] wf query "<question>"             # evidence-backed answers

DO THE WORK (code changes land via the normal commit)

END OF SESSION
  → wf log --project <slug> --receipt <receipt-id>   # outcome against delivery
  → wf gate                                          # what needs a human? surface it

The git hook fires on commit (drift capture, 0 tokens for unchanged files); graphify rebuilds on its own plugin (branch switches), not the agent's attention. Everything the agent prints is either manifest-shaped or a gate report — you never reverse-engineer state from prose.

Why procedures in files, not prompts in code: the skill file versions with the harness (a changed workflow = a changed file = the next session reads the new contract), and it's readable by any agent harness with zero integration — that's the "universal loader" row above.

How skills reach your harness

Surface	Mechanism	Who gets it
Universal loader (wf skill <name>)	prints the procedure on demand	every harness — no skill support required
Native skill folders	wf harness install copies system/skills/<name>/SKILL.md into Claude Code (.claude/skills/) and opencode (.opencode/skill/)	only those two

The content is identical everywhere — the same always-on block (in AGENTS.md/CLAUDE.md) points at wf skill <name>, so harnesses without native skill folders get the same procedures on demand.

Sync contract

.opencode/skill/ and .claude/skills/ copies are generated — wf harness install writes them from system/skills/ (the canonical source). If a skill file changes, re-run install (use --force to rewrite existing copies) and commit both together. Committed copies exist so fresh installs carry the current procedures; they are not hand-maintained.

Which commands have no skill (and why)

Command-wrapper skills (context, lint, query, a vault refresh) were deliberately removed: the CLI itself + its --help + the docs pages are the contract for command-shaped work, and duplicating commands as prose created drift (the deleted refresh skill advertised itself in three docs after its code had been removed). Skills exist only where judgment-based multi-step procedures benefit from an in-context read: the two promotion/ingest workflows carry anti-loop rules and gate order that prose explains better than flags do.

Links

• How It Works: The Compiler — the ingest story in depth
• How It Works: The Compounding Loop — the promotion narrative

---

Next: Configuration