Imported from magicpro97/vibeflow (
.vibeflow/skills/vf/SKILL.md). Install upstream withnpx skills add magicpro97/vibeflow --skill vf. Copyright stays with the author.
Driving work through VibeFlow (vf)
VibeFlow is a local-first orchestrator for AI coding agents (Claude Code, Codex,
Copilot CLI). The golden rule: drive the task through vf, do not free-hand it.
vf already encodes the confidence gate (nothing is "done" < 1.0 with evidence),
the guardrail hooks, the skills store, and the parallel work-unit orchestrator.
Re-implementing those by hand (manual codex exec, hand-rolled review, gh pr create
loops) is the anti-pattern this skill exists to stop.
This SKILL.md is the slim index. Load a reference file only when the task needs it (progressive disclosure) — do not carry the full detail in every turn.
When to use
Use when a task involves VibeFlow workflow setup, orchestration, agent dispatch, or gated verification.
When not to use
Do not use this skill to bypass the SPEC-FIRST gate or replace evidence with a self-reported completion claim.
Steps
- Apply the SPEC-FIRST gate, choose the matching flow, load only needed references, then run the flow.
Verification
vf verifyexits 0 after typecheck, lint, tests, evidence, scope, and normative gates pass.vf skills validatepasses for canonical project skills.
0. SPEC-FIRST GATE (before any writing/dispatching command)
vf init, vf run --yes, and vf orchestrate --yes all WRITE or DISPATCH. Never
run them blind. First reflect the task back as a short spec and get confirmation.
Ask only the questions whose answers you don't already have:
0.1 Spec questions
- Goal — one sentence: what does "done" look like? (becomes the workflow goal)
- Scope — which files/dirs may change? what is off-limits? (drives per-unit scope)
- Engine — claude, codex, or copilot? (default: copilot). Cheap mechanical work → codex.
- Risk class — docs | simple-code | feature | architecture | security | deploy (sets the confidence band; default 1.0 for code).
- Parallel? — one concern → a single unit; several independent slices → one work unit each. Overlapping file scopes are serialised automatically.
- Real run or preview? — default everything to a DRY run first, confirm, then
--yes.
State the spec back in 3-5 lines, ask the open questions, WAIT for the answer, THEN act.
Skip the gate only for read-only commands (vf doctor, vf units status, vf verify,
vf skills list, any --dry-run).
1. Pick the flow
| The user wants… | Flow | Detail |
|---|---|---|
| set up a repo for AI agents | Flow A — init | references/flows.md |
| "here's a spec/issue, implement it" | Flow B — spec → task | references/flows.md |
| several independent changes in parallel | Flow C — workflow | references/flows.md |
| "is it done / ship it" | Flow D — verify & ship | references/flows.md |
bare /vf (no arguments) |
Grill from context | references/grill.md |
Always start with vf doctor --probe if you have not confirmed an engine is ready
this session — a dispatch against a cold engine fails the creation gate.
Bare /vf — grill first
If the user typed /vf with NO task, do not guess and dispatch. Load
references/grill.md: read the recent chat context (last messages, open files,
errors), infer 2-3 likely intents, then relentlessly grill the user through the
SPEC-FIRST questions until the spec is concrete — then map it to a Flow above.
(Interview technique credited to mattpocock/grill-me.)
2. References (load on demand)
references/flows.md— full Flow A-D playbooks (init / spec→task / workflow / verify-and-ship), every flag explained.references/grill.md— bare/vfcontext-grill protocol: infer intents, interview to a spec, route to a Flow.references/hooks.md— guardrail hooks: arming, the live PreToolUse gate, per-engine block-vs-detect semantics.references/pitfalls.md— the anti-patterns learned the hard way; read before improvising.
3. Skills and external docs (before inventing steps)
vf skills list/vf skills search <task>— find a verified skill before hand-rolling.vf skills resolve— report which skill needs are satisfied locally vs. on demand.vf discover docs <lib> --yes— pull external library docs via Context7 (network needs--yes).vf discover skills "<task>" --yes— find an external skill to import.vf skills validate— validate the skill store against the Anthropic format.vf skills sync --engine <name>— mirror.vibeflow/skillsinto an engine's skill dir.
4. Verification (prove it worked)
- After init:
vf doctor(engine ready + hooks armed) and the generated files exist. - After a dispatch/workflow:
vf verifyexits 0 (all gates green) andvf units statusshows the units done at confidence 1.0 with recorded evidence. - When the work edited repo code directly (not via a vf dispatch): the
repo's own gate checklist still applies in full —
bun run fix,bun run check,bun run build, e2e on UI diffs,refresh-normative-proofs, and pre-push review evidence. Load.vibeflow/skills/bun-typescript-conventions§Verification; on this repobun run checkIS the CIcheckjob, so skipping it is how "every change fails CI" happens. - Validate this skill itself:
vf skills validate.
See references/flows.md §Flow D and references/pitfalls.md for the full verify loop.
Powered by VibeFlow.