REQUIREMENT TO PR
dev-skills
From requirement to a verified, mergeable PR: a set of Skills for coding agents.
First, give the repository a verification environment the agent can start, operate and observe on its own. After that, one agent takes each requirement from start to finish and proves the result in the running app, and a model from another family verifies it independently. You step in at three points: confirm the decisions, confirm the acceptance criteria, and read the decision list before you merge.
Description checked on · View the source revision 2254db6
How a requirement gets done
Each repository is prepared once. After that, every requirement goes through definition and delivery. Gaps a run exposes in the repository are listed in the PR, and you decide whether to address them in the repository or in the Skills.
A Readiness · repo-readiness
Let agents verify on their own. You: once per repository, answer what the repository can't tell it.
B Define · core-grill
Ask what only you can decide. You: answer the questions and confirm the decision summary once.
B Define · core-spec
Write down decisions and acceptance criteria. You: confirm spec.md and verify.md once.
C Deliver · deliver
One agent, all the way to a mergeable PR. You: read the decision list before merging, then merge.
A few rules hold throughout. Every decision a person has to make, and every acceptance requirement, is settled at the start. At the end, a model from another family verifies against the acceptance criteria. In between, one owner does the work. Irreversible operations stay with you: merging, force-pushing a shared branch, deleting shared data, sending external messages, changing a shared environment.
A Readiness / Let agents verify on their own
repo-readiness
Done once per repository. Sets up the project's own verification Skill: agents can start, operate and observe the app from its own worktree, with feature maps of numbered criteria, scenario scripts, smoke tests, and a structure check in CI. Whatever the repository can't answer, it asks you in one batch.
What you do: once per repository, answer what the repository can't tell it.
/dev-skills:repo-readiness The next requirement is <link or text>B Define / Ask what only you can decide
core-grill
At the start of a requirement, asks in rounds about the decisions that change what users see: interactions, copy, defaults, entry points, whether data can be lost. Each question comes with options, consequences and a recommendation. The agent settles implementation details itself and records them as defaults. It ends with a decision summary.
What you do: answer the questions and confirm the decision summary once.
/dev-skills:core-grill The requirement is <link or text>; put the requirement documents in <requirement directory>/B Define / Write down decisions and acceptance criteria
core-spec
Turns the decision summary into spec.md (decisions and constraints) and verify.md (acceptance requirements), and has a model from another family check them clause by clause. Once you confirm, both files are frozen and committed to the feature branch, and a draft PR is opened. It can also produce just spec.md.
What you do: confirm spec.md and verify.md once.
/dev-skills:core-spec Write spec.md and verify.md from the decision summaryC Deliver / One agent, all the way to a mergeable PR
deliver
From the frozen spec and verify, one owner implements milestone by milestone and runs the scenarios in the running app. Before a decision, it consults a model from another family, then makes the call and lists it at the top of the PR. Finally, a model from another family verifies the final code independently, and the owner works through CI and review until the PR can be merged.
What you do: read the decision list before merging, then merge.
/dev-skills:deliver Take over <draft PR link>Also used in the workflow
Reviews, PR reading and reports in the workflow use these three Skills. Each also works on its own.
review-rules · Review against clear criteria
Five criteria for reviewing code and designs, re-checking review comments and judging fixes: correct and complete, the minimal necessary change, complexity that matches the benefit, enough extensibility, and a clear division of responsibilities.
Read the SKILL.md ↗/dev-skills:review-rules Re-check this review comment: what's the basis, and what's the smallest fix?mr-for-human · Read the code an agent wrote
Turns a PR into a reading guide for people: first the core conclusions and what needs your judgment, then the design, the execution logic in pseudocode and the runtime constraints, and finally where to find the code.
Read the SKILL.md ↗/dev-skills:mr-for-human Walk me through this PR: <link>explain-as-fool · Explain it plainly
Explains a topic to someone who knows nothing about it, directly and without rhetoric. Reports that other Skills write for people follow its writing rules.
Read the SKILL.md ↗/dev-skills:explain-as-fool What is a thread pool?
Use them on their own
Not part of the delivery workflow. Call them for the task at hand.
design-for-review · A design people can review
Turns requirements and design material into a self-contained technical review document: the problem, the overall design, the scope of the change and the trade-offs that need a decision.
Read the SKILL.md ↗/dev-skills:design-for-review Turn the confirmed design into a review documentplan-for-agents · Turn decisions into a plan
Creates, revises or checks a complete plan for agents to execute, from requirements and confirmed decisions down to steps, outputs and acceptance evidence.
Read the SKILL.md ↗/dev-skills:plan-for-agents Write a self-contained plan from the confirmed specagent-prompt-rules · Rules for writing to agents
Rules for writing and reviewing prompts for agents, multi-agent pipelines and SKILL.md files, each linked to the Anthropic or OpenAI source it rests on.
Read the SKILL.md ↗/dev-skills:agent-prompt-rules Review this SKILL.md against the rulesrecon-to-contract · Turn research into a contract
Turns a comparison of two or more external references into one executable contract with evidence, decisions and acceptance criteria.
Read the SKILL.md ↗/dev-skills:recon-to-contract Compare these three open-source implementations and produce a migration contract
Install and call
Installed as a plugin, the Skills are namespaced: /dev-skills:core-grill in MCode and Claude Code, $dev-skills:core-grill in Codex. Linked from a clone, they have no namespace: /core-grill or $core-grill. Use one or the other, not both.
In Claude Code and Codex, the agent can also call core-spec on its own once a discussion is finished; the other Skills are manual-only. MCode doesn't read that setting yet, so calling Skills by name is the safer choice.
Before you start
You need CLIs for models from two different families, installed and logged in; by default, Claude Code and Codex. The gap check and the final verification must come from the other family. With only one family available, the workflow stops at that step instead of checking with the same family.
You also need Node.js 24, git, and gh or glab for PRs, plus an app the agent can start, operate and observe from a worktree, which is what repo-readiness sets up. Plugin installs cover MCode, Claude Code and Codex; end-to-end tests have only been run in Claude Code and Codex.