Mastermind Build

MasterMind — Build

The single "do it the genius way, end to end" workflow. It runs the loop from ~/.mastermind/engineering/core/agent-loop.md, pulling in the specialist agents and the field pack at each phase. Task: $ARGUMENTS.

Scale effort to the task (~/.mastermind/engineering/core/principles.md): a trivial change skips straight to implement+verify; a foundation gets the full loop. Don't perform ceremony the task doesn't warrant.

The loop

1. Understand — restate the real problem and its scope/lifespan; do the asked task and nothing more

(~/.mastermind/engineering/core/rigor.md → Stay in scope). Read the relevant existing code and conventions first (delegate wide reading to a subagent to protect context); learn the stack if unfamiliar (~/.mastermind/engineering/core/agent-loop.md → Learn the stack first). Match the codebase.

1. Design (non-trivial only) — invoke the architect agent to produce the blueprint:

module/component boundaries + interfaces, state model, data flow, key types, edge-case list. For a multi-file or unfamiliar change, write the design/spec down before coding. Skip for a clear one-file fix.

1. Decide the stack — apply the active field's stack-defaults.md. Choose the simplest thing that

fully works; deviate only for a stated reason. Consult mentors.md if a call is contested.

1. Implement to rigor (~/.mastermind/engineering/core/rigor.md) — build against the design. Handle the unhappy paths

(null/empty/loading/error/zero/one/many/offline/unauthorized/malformed). Types honest, no lazy placeholders, no dead code. Single-purpose units. Style like the surrounding code.

1. Verify — close the loop (~/.mastermind/engineering/core/agent-loop.md) — give the work a check that returns pass/fail

and run it: typecheck + lint + tests + build, and for UI, exercise the actual flow / screenshot. Show the evidence; never assert success. If you can't verify it, it isn't done. Fix root causes.

1. Adversarial review — invoke the code-reviewer agent on the diff in a fresh context. Fix

real must-fix findings (correctness/security/a11y); treat gap-hunting nits as optional to avoid over-engineering. Re-verify after fixes.

1. Capture & report — run the mastermind-levelup skill (capture) to fold any durable lesson

or correction into the active field's lessons.md. Report honestly in a few lines: what shipped, the evidence it works, and anything deferred. Commit / open a PR only if asked.

Non-negotiables

Correctness, security, accessibility are never traded for speed. Speed is the reward for rigor. If an approach is wrong or unsafe, say so once with the better option (~/.mastermind/engineering/core/rigor.md refuse-list), then proceed.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mehrad-dm
• Source: mehrad-dm/mastermind
• License: MIT
• Homepage: https://mastermind.mehrad.me

Install and usage instructions live in the source repository linked above.