Back to directory

SpecShip

A spec-driven autonomous engineering workflow, packaged as a Kiro Power, that walks an AI coding agent through reverse-engineering existing code, planning, test-driven building, adversarial validation and shipping a pull request.

SpecShip

SpecShip runs AI agents through four phases plus an optional brownfield recon step. Recon reverse-engineers existing code into a repo map, stack, components, APIs, data models and business flows before planning begins; Plan brainstorms, does real web research, writes a sprint contract, pre-generates test cases and produces requirements, design and task documents; Build executes the task list milestone by milestone in parallel batches, following a write-failing-test-then-implement cycle with a hard gate that blocks a milestone until typechecking, tests and the build all pass.

Validate runs up to seven adversarial validators in parallel as independent subagents covering code, security, integration, browser, design, alignment and, when a performance requirement exists, load, each producing a typed verdict that is aggregated into a decision to merge, recover or escalate; Ship produces a pull request, changelog and archive with a report of timing, bugs caught and milestone breakdown, or prepares the change locally with push instructions if there is no git remote.

More in Agent skills

See all 581
  • Provenance gate

    github.com

    A deterministic authorization gateway that sits between an AI agent and its tools, tracing where each parameter of a side-effect call came from, and blocking the call or asking a human when the source does not check out.

  • A markdown framework of system rules that restates Isaac Asimov's Four Laws of Robotics as a set of prioritized loss calculations for large-language-model agents to follow, published as a discovery file for agents and crawlers to read.

  • Vitruvius

    github.com

    An engineering research agent that runs a discover, read, synthesize, verify and review loop over technical questions, reading sources directly rather than inferring from titles, and publishing its own scored verifier benchmark, weaknesses included.