_graph.yaml

# pmstack skill graph
#
# Declarative handoff map. Tells skills which prior outputs are useful
# context (not required — skills still work standalone) and what default
# orchestrators (like /sprint) should chain.
#
# This file is read by:
#   - /sprint (chains skills in sequence with confirmation gates)
#   - skill prompts themselves (each skill should glob outputs/ for relevant prior artifacts and use them as context)
#   - decisions-log.md auto-append logic
#
# Edit if you want to change the canonical PM workflow.

version: 0.5.0

# ─────────────────────────────────────────────────────────────────────────────
# Skill graph — for each skill, what prior outputs help it do better work
# ─────────────────────────────────────────────────────────────────────────────
edges:
  # /prd reads /competitive output if present (positioning informs problem framing)
  prd:
    reads_from:
      - skill: competitive
        glob: "outputs/competitive-*.md"
        when: "if a positioning analysis exists for this market, use it for the Target Audience section"

  # /metrics reads the most recent /prd
  metrics:
    reads_from:
      - skill: prd
        glob: "outputs/prd-*.md"
        when: "always — North Star metric should align with the PRD's Success Metrics section"

  # /eval reads /prd + /metrics
  eval:
    reads_from:
      - skill: prd
        glob: "outputs/prd-*.md"
        when: "always — eval failure modes should map to PRD risks"
      - skill: metrics
        glob: "outputs/metrics-*.md"
        when: "always — eval metrics should reuse the measurement framework"

  # /brief can summarize any of the above
  brief:
    reads_from:
      - skill: any
        glob: "outputs/{prd,metrics,eval,competitive,compare}-*"
        when: "the brief should reference the source artifacts so stakeholders can dig in"

  # /compare produces both a plan and an eval YAML; eval YAML feeds /run-eval
  compare:
    feeds_to:
      - run-eval

  # /run-eval is terminal — produces summary.md per run
  run-eval:
    reads_from:
      - skill: eval
        glob: "outputs/eval-*.yaml"
        when: "input — the eval design to execute"
    feeds_to:
      - transcript-review

  # /vibe-test is the layer-1 ritual — read transcripts before writing the eval.
  # Implements Anthropic Steps 0–1 (Demystifying Evals for AI Agents).
  vibe-test:
    reads_from:
      - skill: any
        glob: "outputs/transcripts/**"
        when: "if a transcripts/ folder exists locally, sample from it"
      - skill: prd
        glob: "outputs/prd-*.md"
        when: "PRD provides the 'what should this do' frame"
    feeds_to:
      - eval

  # /transcript-review is the post-run companion — Anthropic Step 6.
  # Walks every failed trial asking: model mistake, grader mistake, or task-spec error?
  transcript-review:
    reads_from:
      - skill: run-eval
        glob: "outputs/eval-runs/*/summary.md"
        when: "input — the run summary to triage"
      - skill: run-eval
        glob: "outputs/eval-runs/*/cases/*.json"
        when: "per-trial detail for the diagnostic walk"
    feeds_to:
      - eval        # rubric + task updates feed back into /eval

  # ───── Default routines (v0.5+) ─────

  # /premortem reads a target PRD and (optionally) same-domain competitive + eval
  premortem:
    reads_from:
      - skill: prd
        glob: "outputs/prd-*.md"
        when: "always — the target PRD; required input"
      - skill: competitive
        glob: "outputs/competitive-*.md"
        when: "if same-topic competitive scan exists, sharpens the failure stories"
      - skill: eval
        glob: "outputs/eval-*.yaml"
        when: "if same-topic eval exists, leading indicators can borrow eval cases"
    feeds_to:
      - prd  # mutates the PRD's Risks section on confirmation

  # /weekly synthesizes the last 7 days of decisions-log + outputs/
  weekly:
    reads_from:
      - skill: any
        glob: "outputs/*"
        when: "filter to last 7 days; group decisions by skill"
      - skill: any
        glob: "decisions-log.md"
        when: "delta on the log is the canonical source for 'decisions made'"

  # /launch-readiness reads the full artifact graph for one feature
  launch-readiness:
    reads_from:
      - skill: prd
        glob: "outputs/prd-<feature>-*.md"
        when: "required — verdict is NO-GO without a matching PRD"
      - skill: metrics
        glob: "outputs/metrics-<feature>-*.md"
      - skill: eval
        glob: "outputs/eval-<feature>-*.yaml"
      - skill: run-eval
        glob: "outputs/eval-runs/*<feature>*/summary.md"
      - skill: premortem
        glob: "outputs/premortem-<feature>-*.md"
      - skill: eval-drift
        glob: "outputs/eval-drift-*.md"
        when: "most recent — flags P0 regression"
      - skill: brief
        glob: "outputs/brief-<feature>-*.md"

  # /lint reads the entire artifact corpus + this graph
  lint:
    reads_from:
      - skill: any
        glob: "outputs/*"
        when: "walks every artifact for graph gaps + cross-artifact drift + staleness"
      - skill: any
        glob: "skills/_graph.yaml"
        when: "this file — defines expected downstream edges"

  # /eval-drift reads prior eval runs to compute a delta
  eval-drift:
    reads_from:
      - skill: eval-self
        glob: "evals/results/*.json"
        when: "self-eval baseline for drift detection"
      - skill: run-eval
        glob: "outputs/eval-runs/*/summary.md"
        when: "user-eval baselines for drift detection"
    feeds_to:
      - brief  # writes a release-blocker brief on regression

  # /onboarding is the entry-point tutorial; reads the example walkthrough
  onboarding:
    reads_from:
      - skill: any
        glob: "examples/walkthrough-code-review/*"
        when: "uses the bundled walkthrough as comparison reference"

# ─────────────────────────────────────────────────────────────────────────────
# Default orchestrators — chains a higher-level command can run
# ─────────────────────────────────────────────────────────────────────────────
orchestrators:
  sprint:
    description: "Customer signal → ready-to-ship-to-stakeholders artifacts"
    chain: [prd, metrics, eval, brief]
    confirm_between_steps: true

  compare-and-run:
    description: "Compare two products, then execute the resulting eval"
    chain: [compare, run-eval]
    confirm_between_steps: true

  onboarding:
    description: "Interactive tutorial for new pmstack users"
    chain: [prd, competitive, premortem, metrics, eval, run-eval, lint, launch-readiness, brief, weekly, eval-drift]
    confirm_between_steps: true

# ─────────────────────────────────────────────────────────────────────────────
# Default routines — five recurring patterns that compose existing skills.
# Each is runnable as a one-shot slash command AND schedulable via /loop.
# ─────────────────────────────────────────────────────────────────────────────
default_routines:
  eval-drift:
    trigger: loop-only
    cadence: 7d
    purpose: "Weekly drift watch over an AI feature; release-blocker on regression"
  premortem:
    trigger: slash-only
    purpose: "Stress-test a draft PRD with 3 failure stories + mutation gate"
  weekly:
    trigger: both
    cadence: 7d
    purpose: "Monday self-snapshot — decisions, open loops, changed-mind"
  launch-readiness:
    trigger: slash-only
    purpose: "Pre-launch GO/NO-GO/CONDITIONAL verdict over the artifact graph"
  lint:
    trigger: both
    cadence: 7d
    purpose: "Workspace audit — graph gaps, cross-artifact drift, stale candidates"

# ─────────────────────────────────────────────────────────────────────────────
# Decision log — every skill that ships an artifact should append here
# ─────────────────────────────────────────────────────────────────────────────
decision_log:
  path: "decisions-log.md"
  format: |
    - YYYY-MM-DD — <skill>: <topic> — <relative-artifact-path>

  rationale: >
    Skills run statelessly. The decision log gives Claude (and you) one place
    to look up what's already been decided, so a future skill knows the
    PRD already exists, the metrics are already locked in, etc.