baseline by friedbotstudio/ docs
friedbotstudio/baseline

Skills

Core skills

The baseline ships 42 skills across thirteen categories. Each runs in main context, where conversational visibility is preserved. The single subagent is a separate mechanism.

§ I

Artifact drafting (4)

Each ships a template.md that the skill fills in. Artifact skills produce the durable workflow documents that downstream phases consume.

  • intake. Phase 1. Captures a new request: problem, desired outcome, constraints, testable acceptance criteria.
  • brd. Cross-functional Business Requirements Document, written when intake is not enough.
  • spec. Phase 4. Diagram-driven technical spec with C4, UML, and a dependency graph in PlantUML.
  • rca. Out-of-band postmortem for incidents and outages. Often precedes a bugfix intake.

§ II

Workflow phases (10)

Auto-invocable. The orchestrator chains them in order, yielding at consent gates.

  • triage. Reads workflows.jsonl, classifies the request by reading each track's selector_hints, and presents the picked track plus alternates via AskUserQuestion. On confirmation, materialises the chosen track's DAG into the TaskList and writes workflow.json.
  • scout. Phase 2. Maps the relevant slice of the codebase.
  • research. Phase 3. Surfaces 2 to 4 candidate solution approaches with concrete tradeoffs.
  • tdd. Phase 6 (solo). Decides scenario recipe and implementation contract; runs scenario then implement then verify then drift-check. Drift-check compares numbered ACs and Design calls rows in the approved spec against the implementation diff and yields the workflow if any item is unresolved (no auto-loop).
  • simplify. Phase 7. Mechanical cleanup pass over the branch diff.
  • security. Phase 8. OWASP-aligned review of pending changes.
  • integrate. Phase 9. Runs the full test suite and stamps the binding verify verdict.
  • document. Phase 10. Routes technical writing through the documentation, technical-tutorials, and prose skills.
  • archive. Phase 10.5. Moves slug artifacts to docs/archive/<date>/<slug>/.
  • commit. Phase 11. Stages and creates the commit, gated by /grant-commit. After the commit lands, if workflow.json declares source_backlog_keys, the commit step invokes sweep.mjs stamp-closure to mark each named backlog entry status: picked-up with superseded-at: <today>; the next /memory-flush auto-close sweep deletes the stamped blocks. Pushes to protected branches need /grant-push (Article VII).

§ III

Phase workers (5)

Execute pre-decided recipes. Each mandatorily invokes a sub-skill so the engineering rules travel with every code-generation step.

  • scenario. Writes failing tests from a recipe. Invokes code-structure.
  • implement. Writes production code that turns scenarios green. Invokes code-structure; uses context7 for any third-party API.
  • verify. Mechanical worker. Runs the test command and stamps the official PASS or FAIL verdict.
  • prose. Drafts or revises English prose for any brief. Always invokes humanizer; conditionally copywriting, documentation, or technical-tutorials.
  • design-ui. Orchestrates impeccable for UI design tasks: classifies intent, picks an impeccable recipe, runs it under a 3-iteration loop cap, persists state at .claude/state/design/<slug>.json. Writes no product code itself. Per Article X.2, every design task in a workflow phase routes through this skill.

§ IV

Spec helpers (4)

Validate the spec artifact before approval and render its diagrams for review.

  • spec-lint. Preflight check before save. Validates PlantUML syntax, required diagram presence, AC-to-sequence traceability.
  • spec-render. User-only. Renders PlantUML diagrams to PNG.
  • spec-diagram-review. Cross-consistency review. Verifies C4 components appear in the dependency graph and so on.
  • spec-traceability-review. Verifies every spec AC traces to an upstream AC in the intake or BRD.

§ V

Orchestration (3)

Drive the workflow end-to-end, plan the swarm path, and dispatch the worker waves.

  • harness. Graph executor. Loads the chosen track, walks the DAG, resolves selector nodes by evaluating preconditions in declaration order, expands sub-tracks inline, dispatches can_parallel clusters concurrently, and yields at consent gates. Runs the one-shot pre-§18 workflow.json migrator at preflight.
  • swarm-plan. Decomposes an approved spec into a dependency-ordered runtime sub-track that the harness reloads as an overlay on the canonical workflows.jsonl. Output lives at .claude/state/swarm/<slug>.jsonl.
  • swarm-dispatch. Executes a swarm plan wave by wave with worktree isolation.

§ VI

Memory (1)

One curation skill bridges the auto-extracted inbox and the canonical files.

  • memory-flush. Reviews auto-extracted candidates in _pending.md and commits keepers to canonical files.

§ VII

The first tool the baseline reaches for on any code-navigation question, in any language. Keyword search keeps landing on the wrong flow when the same name shows up in unrelated code; code-browser follows the actual import graph instead, so you get the right answer the first time. It is preferred over the Explore agent and global grep for structural questions like where data comes from or what wraps a component.

  • code-browser. Its primary path is a language-agnostic universal walk: start at an entry point (a page, route, handler, or command), follow imports to the IO boundary (a network call, a query, a file read), and you have your answer. It works the same in any language. On JS/TS repos, optional accelerators make it instant: discover.mjs maps the project layout once, then walk.mjs runs deterministically and returns flat byHook, byService, byApiCall, and byComponent indexes. It falls back to Explore or grep only when a repo has no resolvable structure or the walk hits a dead end; full-text search and type lookups stay with grep. Read-only.

§ VIII

Phase helpers (1)

Invoked by entry phases as a Step 0.5 gate. Never on user-direct invocation; the gate decides whether the helper fires based on workflow.json.

  • brainstorm. PM-mode requirement capture via Socratic dialogue. Fired at Step 0.5 of /intake, /spec, and /tdd when workflow.json → skip_brainstorm is false. Walks four stages: skip-check, gap-analysis, probe-loop (cap 5 iterations), confirm-and-persist. Writes the brief to docs/brief/<slug>.md with six structured fields (actor, trigger, current state, desired state, non-goals, solution-leakage). The Stage 2 discipline-assertor at discipline.mjs structurally forbids solution-shaped tokens in probes — the SKILL.md prose is the documentation; the regex bank is the enforcement. Opt out via /triage --no-brainstorm. See Article X.3.

§ IX

Generators (2)

On-demand. Not a workflow phase, so they never block a commit. A generator produces a structured artifact a project routes wherever it wants.

  • whatsnew. Emits a structured "what's new" fragment for a set of changes to .claude/state/whatsnew/<slug>.json (gitignored, transient). An optional project.json → whatsnew.route_workflow names a per-project routing workflow that consumes the fragment and renders it wherever the project wants: a site page, a release body, a docs page. Absent that knob, the fragment simply waits at the conventional path. Never writes CHANGELOG.md, which @semantic-release/changelog owns at release time. Replaced the former Phase 11.5 changelog skill.
  • standup. Read-only release and backlog recap. Reports the last release, commits since the last tag classified by type with the bump they trigger, the backlog bucketed, and open questions, then recommends what to pick up next. Runs on demand and as a compact line at session start. Reuses the gather helper; the recommendation is reasoned in main context.

§ X

Shared globals (7)

Vendored or globally available. Available to any phase that needs them.

  • claude-automation-recommender. Vendored under Apache 2.0. Recommends Claude Code automations for a codebase.
  • code-structure. Mandatory on every code-generation step. Enforces orchestration / domain / foundation layering.
  • humanizer. Removes signs of AI-generated writing. Mandatorily invoked by prose.
  • documentation. Reference docs, READMEs, runbooks.
  • technical-tutorials. Step-by-step tutorials, quickstarts, code walkthroughs.
  • copywriting. Persuasive register: landing pages, marketing copy.
  • impeccable. Frontend design and iteration. Mandatorily invoked by design-ui.

§ XI

Audit (1)

One drift defender keeps the implementation honest against seed.md and the constitution.

  • audit-baseline. Drift check between this constitution plus seed.md and the implementation. Read-only.

§ XII

Alternate tracks (1)

A stripped-down pipeline for tasks that need no failing-test-driven code change.

  • chore. Stripped-down workflow for tasks with no failing-test-driven code change: documentation edits, governance count bumps, formatting, dependency bumps. Skips scenario and implement; runs edits directly. Archive, grant-commit, and commit always run. Verify runs by default, but a pure-docs chore can skip it when the project marks its test suite behavior-only (project.json test.kind), so documentation work is not blocked by a code test that cannot apply to it, while real code changes still get verified.