The workflow

CORE

Every change moves through one pipeline. PRD says what changes and why. In spec-driven work, SPEC fixes the exact behavior and contracts, and a human approves it. TECHSPEC grounds the approach in real files. Tasks split it into a dependency graph with tests. Build makes focused changes. Verify collects evidence. Review is a human decision. PR records the rationale and opens the pull request once a human approves the draft. Done is recorded when it has landed. Every stage runs as a skill inside your AI coding agent: maverick-new plans, maverick-build builds, maverick-review verifies and reviews, and maverick-ship drafts the PR. See The Maverick Method for the stage-by-stage playbook and From idea to tasks.

Pick a workflow and click a stage to see what it produces:

Standard — Features, fixes and refactors.

prd

Say what changes — and what does not

The maverick-prd skill reads the grounded code, asks one question at a time, and writes goals, user stories, requirements with scenarios and acceptance criteria.

ARTIFACT PRD.mdOverview · Goals · User Stories · Requirements · Acceptance Criteria · Business Rules · Edge Cases · Non-Goals · Open Questions

YOU DECIDEAnswer the product questions the code cannot answer, then confirm the next stage.

RUN IN YOUR AGENT

skillmaverick-prd

Claude Code/maverick-prd add-orders-pagination

Codex · Cursor · Gemini · Copilot · KiroUse the maverick-prd skill for add-orders-pagination

CLI GATE

$ maverick validate add-orders-pagination --stage prd

Start a new task with maverick new "<idea>": it creates the task and opens your agent at this stage.

Run each stage

Every stage is a skill. In Claude Code, type it as a slash command with the task slug. In Codex, Cursor, Gemini, Copilot, Windsurf, Kiro or OpenCode, ask for it by name: "Use the maverick-spec skill for add-orders-pagination".

Stage	Run in your agent	It writes	Gate (CLI)	You decide
PRD	/maverick-prd <task>	PRD.md	maverick validate <task> --stage prd	answer product questions
SPEC spec-driven	/maverick-spec <task>	SPEC.md	maverick flow run spec <task>	◆ approve the spec
TECHSPEC	/maverick-techspec <task>	TECHSPEC.md	maverick validate <task> --stage techspec	confirm the approach
TASKS	/maverick-tasks <task>	subtasks/	maverick validate <task> --stage tasks	confirm the graph
BUILD	/maverick-build <task>	code + tests	maverick subtask next <task>, maverick verify <task>	OK each next subtask
VERIFY	/maverick-review <task>	EVIDENCE.md	maverick verify, guard, deps	read the evidence
SECURITY strict	/maverick-review <task>	SECURITY-REPORT.md	maverick security <task>	read the findings
REVIEW	/maverick-review <task>	REVIEW.md	maverick flow run ship <task>	◆ record the decision
PR	/maverick-ship <task>	PR-DRAFT.md, the pull request	maverick flow run pr <task>	◆ approve the draft, then merge
DONE	—	TASK.json	maverick done <task>	finish the task

You don't have to remember the table. /maverick-new <idea> chains the planning skills for you and asks before each one, maverick new <task> reopens the agent with the right skill for the current stage, and maverick status names it in each task's next step. See The Maverick Method for each stage in depth.

Workflows

Workflow	Best for	Adds
standard	Features, fixes and refactors. This is the default.	—
spec-driven	APIs, integrations, shared contracts, complex domain rules, or work several people must agree on.	SPEC.md between the PRD and the technical spec, and the spec approval gate.
strict	Security, data, payment, migration or shared-infrastructure work.	A security gate before review and a checkpoint before merge.

maverick new "Add pagination to the orders list"            # standard
maverick new "Expose the orders API to partners" --spec     # spec-driven (SDD)
maverick new "Replace the shared cache layer" --strict      # strict

maverick new picks strict on its own when the idea touches security, data, payments or migrations, and spec-driven when it defines an API, an endpoint, a webhook, an integration or another contract. Pass --workflow standard to skip the spec. Change the default with workflow.default in .maverick/config.json.

Spec-driven development (SDD)

spec-driven puts a behavioral spec in the middle of the pipeline:

PRD ──► SPEC ──► [human approves] ──► TECHSPEC ──► TASKS ──► BUILD ──► VERIFY ──► REVIEW ──► PR ──► DONE
what     exact behavior                how           graph      tests first
and why  and contracts                                          against SPEC

• PRD.md stays the product contract: goals, stories and AC-NN criteria.
• SPEC.md (the maverick-spec skill) turns each criterion into numbered SC-NN WHEN/THEN scenarios, exact contracts (requests, responses, events, signatures), states and data, invariants, errors, and a traceability table from every AC-NN to its scenarios and contracts. maverick validate <task> --stage spec fails when a scenario lacks WHEN/THEN or references an undefined criterion, and, under --strict, when a criterion is not traced.
• The spec flow is the gate: maverick flow run spec <task> checks the sections and the validation, then waits until a human checks - [x] Spec approved in SPEC.md or runs maverick flow approve spec <task> spec-approval.
• TECHSPEC.md implements the contracts exactly and cites the scenarios in its Test Plan (IT-01 → AC-01 / SC-01). Build writes the scenario tests first; review checks every scenario has a passing test and every contract matches SPEC.md.

A behavior change after approval starts in SPEC.md, not in the code.

Know where each task is

maverick status

status shows each task's stage and names the skill to run next; maverick new <slug> reopens the agent with it. Under the skills, guard checks the diff against the TECHSPEC File References and the task Files, and review <task> --diff writes REVIEW.md from the PRD acceptance criteria and the tasks.

Automate the loop

When the stages are clear, maverick flow run ship <task> runs contract-driven gates — checks, commands, artifacts and human approvals — until the task is done, stalled or waiting for a person. Add --agent to let a trusted, permission-limited agent fix blocking issues between runs. See the CLI reference.