Real-User QA Planner

Plan QA as journeys real people walk, not test cases that accumulate. This skill owns the project's living QA docs — one committed tree (<qa-docs-path>, default docs/qa/) that every round appends to — and plans the persona-driven sessions qa-execution runs.

Two rules anchor everything:

1. Living docs, not round artifacts. All durable QA knowledge lives in the one committed tree; rounds append to it (structure, durability, and anti-patterns: references/qa-docs-layout.md).
2. Sessions, not cases. The atomic planning unit is the session charter (persona + journey + tour + time-box), derived from journey flowcharts. Coverage means "every planned journey was walked by a persona this cycle" — a session ledger, never a per-case count.

Choose smoke/targeted/full scope from the requested change before following the relevant steps. Reuse current personas, journeys, charters, and evidence; read only the reference sections whose schema or procedure is needed. A targeted update does not re-bootstrap or audit the entire QA tree.

Required Inputs

• qa-docs-path (optional): root of the living tree; defaults to docs/qa at the repo root — a durable, committed location, never a temp dir. Honor an explicitly supplied path; use a repository-owned durable tree by default. Ask only if the intended output location is unresolved.

Procedures

Step 1 — Resolve or bootstrap the tree. Read references/qa-docs-layout.md (canonical tree, gitignore block, bootstrap procedure, adoption procedure for scattered legacy artifacts). Resolve <qa-docs-path>. If the tree exists, read its README.md and search the affected scenarios/ and related open bugs/ first, and build every decision below on that state; when the branch just merged parallel QA work, reconcile before planning — two files describing one behavior or one symptom fold into the older id (merge verdict fields by last_report recency, update references, delete the duplicate, record the fold in the cycle's report). If the tree does not exist, bootstrap it per the layout reference — directory tree, seeded templates/, and the gitignore block. Adopt legacy QA artifacts only when that migration is in scope; preserve historical evidence during ordinary targeted updates.

Step 2 — Establish project personas. Read references/personas.md (seed catalog + derivation rules). Personas are durable instance data in <qa-docs-path>/personas.md: update them only when the product's audience changed; if absent, define the persona needed for the in-scope journey; expand the catalog when audience coverage requires it.

Step 3 — Map journeys as flows (before any scenario). Read references/journeys-and-flows.md (journey anatomy, Mermaid mapping, flows-before-matrix). Scope the mapping: a branch/PR cycle covers every user-visible change in the diff; a release cycle covers the product's high-value journeys. For each, write or update <qa-docs-path>/journeys/J-<slug>.md — the YAML journey map plus a Mermaid flowchart from entry → actions → branch points → side effects → the true end state, with at least one abandonment path. Map the flow first; the scenario comes from it.

Step 4 — Derive scenarios into the tracker. Read references/state-schema.md (fields, enums, id minting — exact) and references/taxonomy.md (the five coverage dimensions). Walk each flowchart and derive scenarios: one scenarios/<AREA>-<slug>.md file per scenario with a content-addressed id, updated in place, overlaps recorded in the overlaps field. Use the taxonomy dimensions relevant to the changed journey; a full release plan can sweep all five. Scenario files are planning output — qa_status stays untested until qa-execution runs them.

Step 5 — Plan session charters. Read references/session-charters.md (charter anatomy, cadence tiers, the coverage inversion). Pick the cadence tier (smoke / targeted / full / sanity); the tier picks the journeys. Write one charter per session to <qa-docs-path>/charters/CH-<slug>.md from <qa-docs-path>/templates/charter.md (seed: assets/charter-template.md), preserving its headings — mission, persona, journey, exactly one tour, time-box, must-try guidance — ordered by risk: highest-impact journey × highest-blast-radius tour first. Reuse an existing charter whose mission still fits before writing a sibling.

Step 6 — Register bugs. Read references/bug-registry.md (id minting, dedup, the five-tier user-impact rubric — the canonical severity model for both skills). Dedup before filing: search <qa-docs-path>/bugs/ for the symptom and update the existing file rather than duplicating — a re-found bug is history worth keeping on one id. Only a genuinely new symptom mints a new content-addressed BUG-<YYYYMMDD>-<slug> id; write it from <qa-docs-path>/templates/bug.md (seed: assets/bug-template.md), preserving its headings, and link the id into the affected scenario files' bug_ids.

Step 7 — Validate cycle completeness. For a new/full cycle, verify the following planning contracts. A targeted cycle reuses unchanged maps/charters and checks only its affected entries; record gaps honestly rather than padding:

• every in-scope journey has a flowchart with a true end state and ≥1 abandonment path;
• every in-scope journey has ≥1 charter with an assigned persona;
• every in-scope scenario file has a content-addressed id, a linked journey, and a qa_status reflecting reality;
• every open bug has a registry file and appears in ≥1 scenario's bug_ids;
• the five taxonomy dimensions were considered per journey — a skipped one is recorded with reasoning.

The completeness bar is "every journey walked by a persona", a session ledger — never a per-case count. Case accumulation is the failure mode this skill exists to prevent.

When a journey grows stable or regression-prone enough to deserve an automated E2E spec, read the applicable section of references/automation-backlog.md, then record the intent as one file in <qa-docs-path>/automation-backlog/ — one backlog, never automation fields on individual scenarios or charters.

Companion Skills

• qa-execution — runs the sessions this skill plans and writes results back into the same tree (statuses, bugs, reports). The living tree is the contract between the two.
• agent-output-audit — owns CI verification gates, AI test-hygiene scans, and task-status reconciliation. Route technical integration/security/performance/load suites there or to dedicated tooling; record the routing decision, don't absorb the work.

Error Handling

• A scenario file's frontmatter won't parse (missing delimiter, unknown field, nested value): repair it and report what was repaired before any downstream step — every step depends on a loadable tracker.
• Two files describe one behavior or one symptom under different slugs (typical after merging parallel QA branches): run the Step 1 fold before any downstream step plans on top of the duplicates.
• A branch cycle's diff has no user-visible change: say so and stop; there is nothing to dogfood. Do not invent scenarios to fill a cycle.
• <qa-docs-path> can't be created (permissions, read-only checkout): surface the error and stop — never fall back to a temp directory.