AI Workflow MVP Scope
Scope an AI workflow MVP to one user, one workflow, one metric, with human-in-the-loop and failure design on every model step. Use when designing an AI feature, writing MVP scope, or checking an agent workflow before build.
Copy the file into your agent runtime. Cursor loads personal skills from ~/.cursor/skills/ai-workflow-mvp-scope/SKILL.md and project skills from .cursor/skills/ai-workflow-mvp-scope/SKILL.md. Claude Code looks in .claude/skills/ai-workflow-mvp-scope/SKILL.md.
AI Workflow MVP Scope
When to use
An "AI product" is about to sprawl. Triggers: scope the MVP, design this agent workflow, one user one workflow, HITL for this step.
What it does
Reframes the problem, forces one user / one workflow / one metric, designs each step (human / script / model), and refuses to start build without KPIs, cost/latency envelopes, and failure paths. Drawn from ARC's Nova operating procedure, generalized.
Steps
- Reframe: primary user, job, decision supported, what "better" means (faster / more accurate / cheaper / less fatigue). If this is fuzzy, stop and ask.
- Concept brief: one user, one workflow, one metric. Pre vs post. Classify every step.
- MVP in/out: what is built, what stays manual, minimum data and integrations, go-live acceptance.
- For every model step answer: task (classify/extract/summarize/generate/decide), input/output shape, accuracy range, how accuracy is measured, low-confidence behavior, unavailable behavior, wrong behavior, who reviews when, logging/retention, cost per run, latency envelope, whether the data is allowed to leave.
- KPIs: business, system, adoption — each with definition, target, measurement, review cadence, escalation.
- Rollout: pilot population, onboarding, feedback, rollback, phase-2 criteria.
Output
Problem reframe, concept brief, MVP in/out, per-step design table, KPI table, rollout, open questions.
Guardrails
Do not design a platform when an MVP would do. Do not let model choice drive the workflow. Do not skip KPIs "until after launch." Do not treat HITL as a later add-on. If data classification is unknown, the model step is blocked, not "we'll be careful."