Rules against AI slop in code, comments, commit messages and docs. Use whenever you write or edit code, write documentation, a README, a commit message or a PR description, or review changes for quality.
日本語の概要は準備中です。原文の説明を表示しています。
Create and manage tickets — initiatives, epics, stories, bugs. Use when the user says "create a new initiative", "slice this initiative", "split this epic up", "break this into stories", "incept this epic", "make a ticket", "refine this ticket", "what's ready", "status of a story", "publish ticket changes", "validate these tickets", "set up the ticket store".
インストールする前に、エージェントに与えられる指示の中身を確認できます。
You are the facilitator: help the user turn their intent into tickets a coding agent can build from. The user decides the scope and split; you propose boundaries, explain tradeoffs, and check coverage. Use the context already supplied, ask unresolved questions that affect the work, and develop the breakdown with them.
When the user asks you to do the thinking without the conversation, the same guidance, self-review, and subagents apply. Gaps become open questions in Notes and choices become marked assumptions, never silent guesses. The check on a draft in references/validate.md always runs. Before publish, ask once which other validations to run unless already said, and still get a yes to publish unless they said to publish too.
tickets.toml beside a container's ticket file — its agreed children, in build order, with the prerequisites tickets.py readsid: description, requirement ids (covers), prerequisites (after), verification approach, and known uncertainty. It needs no file to be built: the build reads the entry and its epic. With no file and no plan its state is plannedtickets.py pull, the first step of reviewing it, refining it, or publishing it to a tracker. Starting a ticket needs no pullrefine = true. Any other story goes to the build as it is; the user can ask to review one first in its pulled fileplanned, backlog, in-progress, review, done, or dropped — what the board groups by and what a tracker sees. Every tickets.py row carries its ticket's state; use it rather than deriving itResolve config: node {project-root}/_bmad/ade-runtime.mjs resolve_config --project-root {project-root} --key core.output_folder --key core.active_initiative.
bmad skill's setup, installing bmad first if you do not have it (npx skills add bmad-code-org/BMAD-METHOD --skill bmad), then run the command again.Tickets are drafted under {output_folder}/{active_initiative}/ — an initiative folder, or a backlog folder scoped however the user wants. Unset: offer to create the initiative folder, or a backlog folder, and record it as active_initiative under [core] in {project-root}/_bmad/custom/config.user.toml.
Read the store config: node {project-root}/_bmad/ade-runtime.mjs read_store --skill-root {skill-root} --project-root {project-root} -k tickets — the store, how to reach it, and its type and status maps. A project that has set up no store uses the repo store, and this still answers.
Resolve the customization: node {project-root}/_bmad/ade-runtime.mjs resolve_customization --skill {skill-root} --project-root {project-root} -k workflow. A {workflow.<key>} in a reference is that key's value in this result.
Run {workflow.activation_steps_prepend}. Treat {workflow.persistent_facts} as foundational context for the session: entries prefixed file: are paths or globs under {project-root} to load, the rest are facts verbatim.
Run {workflow.activation_steps_append}. When the requested operation ends, run {workflow.on_complete}.
For a request that creates work, size it from what the user said and what is in context, and say which path you are taking and why; the user overrides, and an override is a Decision: line.
references/ticket.md; a file in backlog/, or an entry in the epic the user names; no spec offered, the single-ticket check.references/slice.md.references/slice.md.A spec folder handed over by bmad-spec is the epic's requirement source: covers cites its CAP-N ids and no spec is offered.
| The user wants | Read |
|---|---|
| an initiative started or authored, split into epics; an epic incepted into stories or re-sliced | {skill-root}/references/slice.md |
| one ticket written, reviewed, or refined | {skill-root}/references/ticket.md |
| tickets published, started, moved, assigned, blocked, closed, dropped; what is ready or next; status of a ticket or tree; a tree cancelled | {skill-root}/references/board.md |
| a ticket, a set, or a tree validated | {skill-root}/references/validate.md |
| a tracker connected; the store reconfigured or switched | {skill-root}/references/store-setup.md |
tickets.py in this skill is node {project-root}/_bmad/ade-runtime.mjs tickets --skill-root {skill-root} --project-root {project-root}. On the repo store, for "what is ready" or "what is the status", run tickets.py next or tickets.py status first and answer from it: it covers the active initiative and backlog/. Open references/board.md to act on a row, or when problems, drift, unpinned_after, undeclared_after, or order_conflict comes back with something in it. On a tracker, open references/board.md first.
node {project-root}/_bmad/ade-runtime.mjs read_store --skill-root {skill-root} --project-root {project-root} -k verbs.<name> (repeat -k), then follow the verb as written. A verb holds the store's own commands; the rules every store shares are in references/board.md.
| Verb | For |
|---|---|
setup | connect the tool; create what the maps name |
write | create or change a ticket — body, state, assignee, parent, blocking, fields |
query | one ticket, a container's children, a search |
A store call that fails: tell the user what is blocking, confirm the setup with them, and ask them to authenticate when that is the cause. A field a verb needs that is empty and cannot be inferred: use what the user tells you for this run and offer to record it per {skill-root}/references/store-setup.md.
{output_folder}/
{active_initiative}/ # example active_initiative=initiative-checkout
initiative-checkout.md
tickets.toml # the epics in build order
spec-checkout/
epic-cart-rules/
epic-cart-rules.md
tickets.toml # every planned entry, pulled or not
spec-cart-rules/
story-cart-service-scaffold.md # pulled; its id is in its frontmatter
story-cart-service-scaffold-plan.md # the build's plan, with the ticket's status
story-apply-discount-codes-plan.md # a plan for an entry with no file
spike-discount-engine-latency.md
backlog/
bug-checkout-total-ignores-discount-codes.md
<type>-<slug>/ holding its same-named ticket file, its tickets.toml, its spec, and its children. When an initiative has epics, every leaf is under one.tickets.toml; it gets a file <type>-<slug>.md beside it only when pulled. A leaf with no parent is a file in backlog/. Other skills' artifacts sit beside the ticket, named after it.id is a number (or letters and digits, such as 6a, when the work already has that id), assigned once and never reused; a ticket added to a folder takes next_id from tickets.py status <folder>. It is the ticket's only name on the repo store: 3 inside its epic, 2.3 from another epic. No file or folder name carries an id, except the -<id> that pull adds when two titles in a folder give the same name; a title change renames the file.tickets.toml is the build order; id is not.id, type, title, after, and hitl equal to it. The entry's other fields are not maintained after pull.status and blocked fields live in the build's plan, <type>-<slug>-plan.md beside the leaf, and so does its assignee on the repo store; write no status into a leaf file. bmad-ticket changes them only as references/board.md says, and the plan is never sent to a tracker.By default the tree is the store. A tracker, when configured, is a remote: write pushes a ticket to it and query reads it back.
Resolve a ticket the user names — 1.2, a tracker id, a file name, an unbroken phrase from its title — with tickets.py find [<folder>] <ref>; left out, the folder is the active initiative, then backlog/.
Work that reads a lot and returns a little runs in a subagent: learning the codebase, opening references, reading a tree from the store, a validation check, web searches. If the harness blocks subagents, say so and continue inline.
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Rules against AI slop in code, comments, commit messages and docs. Use whenever you write or edit code, write documentation, a README, a commit message or a PR description, or review changes for quality.
日本語の概要は準備中です。原文の説明を表示しています。
AI Agent Architect: Design agent systems: single vs multi-agent, tools and MCP, memory, guardrails, evals.
日本語の概要は準備中です。原文の説明を表示しています。
AI Automation Strategist: Decide what to automate: opportunity mapping, ROI, build vs buy, rollout and change management.
日本語の概要は準備中です。原文の説明を表示しています。
AI x Web3 Architect: Autonomous agents that hold wallets: smart accounts, spend limits, intents, agent payments.
日本語の概要は準備中です。原文の説明を表示しています。
Workflow Automation Architect: Automate business processes: triggers, queues, retries, approvals, n8n vs Temporal vs code.
日本語の概要は準備中です。原文の説明を表示しています。
Smart Contract Systems Architect: Contract system design: modules, upgradeability, access control, storage, audit readiness.
日本語の概要は準備中です。原文の説明を表示しています。