Workflow Skill

Five modes for structured multi-phase work. Match the request, follow that mode's instructions.

Mode Selection

Feature Lifecycle

Phase-gated workflow: DESIGN > PLAN > IMPLEMENT > VALIDATE > RELEASE. Each phase must pass its gate before the next begins.

Phase Routing

If .feature/ exists, check state: python3 ~/.claude/scripts/feature-state.py status. Route to the indicated phase.

If no feature state exists, determine entry from intent:

• "design", "think through", "explore approaches" -> DESIGN
• "plan", "break down", "create tasks" -> PLAN (requires completed design)
• "implement", "execute plan" -> IMPLEMENT (requires completed plan)
• "validate", "quality gates" -> VALIDATE (requires completed implementation)
• "release", "merge", "ship it" -> RELEASE (requires passed validation)
• "end to end", "full lifecycle" -> DESIGN (start from beginning)

Phase References

Load the phase reference, then follow it exactly:

State operations use python3 ~/.claude/scripts/feature-state.py only. Never manipulate state files directly.

Planning

Spec writing, plan creation, interviews, ambiguity triage, and session pause/resume. Planning owns specs and saved plans; execution goes through the process skill's Subagent mode or workflow dispatch.

Sub-mode Routing

For interviews, batch independent questions into frontier rounds. Ask dependent questions sequentially. Include a recommendation per question.

Objective Loop

Iterate-until-verified-done loop. A user states an objective with verifiable done-criteria; each iteration routes one /do cycle, verifies by executing the criteria, and reschedules until verified-done or budget-stop.

Phase 1: SPEC

Gather from the request: objective statement, DONE-CRITERIA (verifiable checks), iteration budget (default 5), NOT-DONE-YET guardrails (what may never be done to satisfy a criterion).

DONE-CRITERIA types: command (preferred -- deterministic command with expected exit code/output) or rubric (only when no mechanical check exists -- frozen at SPEC time, graded by a fresh-context agent).

Phase 2: STATE

Write .objective/<slug>/state.md from references/ol-state-file.md. Wakeups resume from the state file, never conversation memory.

Phase 3: ITERATE

Plan the smallest next step. Route through /do: classify -> route -> dispatch agents -> evaluate. The loop dispatches exclusively through /do -- never edit inline.

Phase 4: VERIFY

Run every done-criterion check. A worker's "passes" claim never substitutes for re-running.

• command: run it, paste exit code and output into iteration log.
• rubric: dispatch fresh-context sub-agent (did NOT produce the work) with artifact + rubric only. Returns PASS/FAIL with cited evidence.

All pass -> final report, STOP. Any unmet -> Phase 5.

Criteria-gaming guard: a criterion may never be satisfied by weakening a hook, gate, test, or safety control. Stop and report the conflict if that is the only visible path.

Phase 5: RESCHEDULE or STOP

All pass: stop. Unmet + iterations remain: update state file, call ScheduleWakeup (delay 270s for active polling, 1200s+ for idle work). Budget exhausted: honest NOT-DONE report with per-criterion status.

Hill Climb

Metric-driven optimization loop. One number moves; everything else stays fixed. Each iteration: hypothesis -> one change -> correctness floor -> re-measure -> accept or revert.

Phase 1: SPEC

One METRIC per loop. Two numbers with a trade-off: promote one to the FLOOR. Load references/hc-domain-playbooks.md for pre-filled SPEC blocks per domain (frame rate, API latency, CI time, bundle size, memory, token cost).

Phase 2: BASELINE

Run MEASURE N times (N >= 5, N >= 10 for wall-clock). Record median and spread. If spread >= target improvement: STOP -- harness too noisy. Report noise sources and offer to stabilize first.

Phase 3: PROFILE

Locate the cost before changing anything. Load references/hc-profiling-tools.md for per-domain tooling. Guessing at hot spots is the dominant failure mode.

Phase 4: HYPOTHESIZE and CHANGE

State one hypothesis targeting the profiled hot spot. Make one change. Run FLOOR commands -- revert immediately if any fail.

Phase 5: MEASURE

Run MEASURE N times. Compare median to baseline. Accept only if delta > variance tolerance. Update ledger (references/hc-ledger.md). If accepted, new baseline.

Phase 6: LOOP or STOP

Target reached: final report. K consecutive non-improving iterations: plateau stop. Budget exhausted: report what worked and what remains.

Ad-Hoc Workflow

For structured multi-phase work that does not fit the four modes above. Identify the workflow from the table, load its reference, follow its phases exactly.

Cost Gate

Ask first: does this need a multi-agent workflow? Skip the workflow when: single-file mechanical edit (use quick), one agent satisfies the request (direct dispatch), lookup/status/count (direct agent). Escalate only when the request has independent subtasks, needs orthogonal verification, or names "comprehensive / thorough / adversarial / tournament."

Composable Patterns

Workflow Catalog

Load the reference for the matched workflow. references/... paths resolve under ${CLAUDE_SKILL_DIR}.

Terminology

"Workflow" is the canonical term. "Pipeline" is the retained legacy alias -- kept for back-compat in routing keys, meta.name exports, and pipeline-index.json. Use "workflow" in new prose; do not rename code identifiers.

Error Handling