mk:autobuild

What This Skill Does

Autonomous build pipeline for green-field products. Runs planner → contract → generator ⇄ evaluator loop with scaffolding density auto-tuned to model tier. For single features, use mk:cook.

When to Use

• Building a green-field product from scratch (kanban app, SaaS platform, retro game maker)
• Multi-hour autonomous builds with no manual handholding
• Any "build me an app" or "create a [product]" request
• NOT for: single features, bug fixes, or refactors (use mk:cook), documentation updates, or explain/review requests

Core Capabilities

• Adaptive density: Auto-selects MINIMAL / FULL / LEAN scaffolding based on model tier via mk:scale-routing, honors MEOWKIT_AUTOBUILD_MODE override
• Generator ⇄ evaluator loop: Developer subagent builds, evaluator verifies against rubrics, iterates up to 3 rounds before escalating
• Budget enforcement: Warns at $30, requires approval at $100, hard-blocks at user-set --budget value
• Run reports: Every run produces a full audit trail at tasks/autobuild-runs/YYMMDD-{slug}/run.md including budget trail, per-step artifacts, and final verdict
• Resumability: Checkpoints at every step — if killed mid-run, --resume {run-id} picks up at the last completed step
• TDD opt-in: --tdd writes the .claude/session-state/tdd-mode sentinel; developer waits on tester before each generator iteration

Usage

/mk:autobuild build a kanban board with drag-and-drop
/mk:autobuild build a todo app --tier full              # Force FULL density
/mk:autobuild build a retro game maker --tier lean       # Force LEAN
/mk:autobuild build a markdown editor --budget 25 --max-iter 2
/mk:autobuild --resume 260501-1450-build-kanban          # Resume interrupted run
/mk:autobuild "build payment system" --tdd               # TDD enforcement

Example Prompt

Build a kanban board with drag-and-drop, swim lanes, and real-time collaboration. Support up to 10 concurrent users. Use a $50 budget cap and max 2 iteration rounds.

Workflow (step-file — one step at a time)

Step	Action
0 — Tier Detection	mk:scale-routing picks MINIMAL / FULL / LEAN; honors MEOWKIT_AUTOBUILD_MODE
1 — Plan	mk:plan-creator --product-level — user stories, not file paths
2 — Contract	mk:sprint-contract (FULL: required, LEAN: optional if ACs < 5, MINIMAL: skip)
3 — Generate	Developer subagent with 4-subphase pattern (Understand → Design → Implement → Verify)
4 — Evaluate	mk:evaluate — active verification against rubrics
5 — Iterate or Ship	PASS → shipper; FAIL → loop to step 3 (max --max-iter rounds, default 3)
6 — Run Report	Append-only audit trail at tasks/autobuild-runs/YYMMDD-{slug}/run.md

Density modes

Mode	Model	Planner	Contract	Iterations	Context Reset
MINIMAL	Haiku	Skip (use --fast)	Skip	Skip	Skip
FULL	Sonnet, Opus 4.5	Required	Required	1-3 rounds	Optional
LEAN	Opus 4.6+	Required	Optional	0-1 rounds	Skip (auto-compact)

Override: --tier flag or MEOWKIT_AUTOBUILD_MODE. Density never bypasses gates.

Hard constraints

• Budget: warns at $30 spent, requires explicit approval at $100, hard-blocks at user-set --budget value
• 6-hour wall-clock timeout — hard limit per Anthropic's observed runs; checkpoints every step for resumability
• Max 3 iteration rounds between generator and evaluator before escalating to human
• Run report mandatory — every run writes full audit trail
• TDD opt-in: --tdd writes .claude/session-state/tdd-mode sentinel; developer waits on tester before each generator iteration

Outputs

Artifact	Location
Run report	tasks/autobuild-runs/YYMMDD-{slug}/run.md
Product plan	tasks/plans/{plan-dir}/plan.md
Sprint contract	tasks/contracts/{date}-{slug}-sprint-N.md
Evaluator verdict	tasks/reviews/{slug}-evalverdict.md
Exit status	PASS | WARN | FAIL | ESCALATED | TIMED_OUT