菜单
Use this skill when interacting with the Kanbanzai system as an agent: before starting implementation work, when assembling context, dispatching or completing tasks, writing commits, contributing knowledge entries, or spawning sub-agents. Activate for any task, not just complex ones.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use this skill whenever you are drafting, reviewing, or discussing a design document, helping a human think through a design problem, presenting alternatives, or preparing a design document for approval. Also activates when asked to make architecture, technology, or data model decisions — those belong in a design document, not in chat.
Use this skill whenever you create, edit, register, or approve any document in a Kanbanzai-managed project, including files in work/design/, work/spec/, work/plan/, work/dev/, work/research/, work/report/, work/review/, work/retro/, or any configured document root. Activate even when document registration is not the primary goal of the task — any file creation in a document root requires registration.
This repository is managed with Kanbanzai. Read this skill at the start of every session, before writing any code or running any searches. Kanbanzai provides MCP tools — next, entity, doc, status, and others — that replace manual grep and file searching for project state and work queue management. The presence of a .kbz/ directory or kanbanzai tools in your tool list confirms you are in a Kanbanzai-managed project and this skill applies.
Use this skill when conducting a plan review after all features in a plan are complete. Covers scope verification, feature completion checks, spec conformance, documentation currency, and retrospective contribution.
Use this skill during any planning conversation: when scoping new work, deciding whether something is one feature or a plan with multiple features, responding to a human request to build something new, or when a conversation might drift into design decisions before scope is established.
Use this skill when conducting a code review for a kanbanzai feature. Provides structured review procedure, per-dimension evaluation guidance, output format, and orchestration steps.
| name | kanbanzai-agents |
| description | Use this skill when interacting with the Kanbanzai system as an agent: before starting implementation work, when assembling context, dispatching or completing tasks, writing commits, contributing knowledge entries, or spawning sub-agents. Activate for any task, not just complex ones. |
This skill defines how agents interact with Kanbanzai: assembling context, dispatching and completing tasks, writing commits, contributing knowledge, and spawning sub-agents.
Before starting work on any task, call next with a task ID to claim it and receive a
context packet:
next(id="<task_id>")
The returned packet is byte-budgeted and prioritised — it contains the task instructions, relevant knowledge entries, and design context trimmed to fit the budget. Read it before doing anything else.
After completing the task, record which knowledge entries were used or found incorrect. This drives auto-confirmation and auto-retirement:
knowledge(action="confirm", id="KE-...")knowledge(action="flag", id="KE-...", reason="...")Call next (without an ID) to see ready tasks.
Call next(id="TASK-...") to atomically claim a task (transitions it from ready to active).
Do the work.
Call finish when done:
finish(
task_id="TASK-...",
summary="What was done",
files_modified=["path/to/file.go"],
verification="What was tested or verified",
knowledge=[...]
)
Every commit must follow this format: <type>(<object-id>): <summary>
| Type | When to use |
|---|---|
feat | New feature behaviour |
fix | Bug fix |
docs | Documentation change only |
test | Test-only change |
refactor | Behaviour-preserving structural improvement |
workflow | Workflow-state-only change (no code) |
decision | Decision-record change |
chore | Small maintenance change with no better category |
Add ! after the type for breaking changes: feat(FEAT-001)!: description
Examples:
feat(FEAT-152): add profile editing API and validationfix(BUG-027): prevent avatar uploads hanging on large filesworkflow(TASK-152.3): mark upload task complete after verificationWhen you discover something reusable, contribute it:
knowledge(
action="contribute",
topic="short-topic-slug",
content="Concise, actionable statement",
scope="<role or project>",
tier=3
)
Tier 3 is session-level (30-day TTL). The system auto-promotes entries to tier 2 based on
usage across sessions. You can also contribute via the knowledge argument in finish.
When spawning sub-agents, include in every delegation message:
next and handoff are the primary skill delivery mechanisms for MCP-connected sub-agents.
Prefer them over manually reciting conventions.
next fails when claiming a task: Another agent already claimed the task. Call next
again (without an ID) to get an updated list of ready tasks.
Tool call failures: Read the error message before retrying. Most failures are deterministic — retrying with the same arguments will produce the same error.
Missing verification summary: finish will succeed without verification, but it
degrades review quality and reduces the value of the knowledge base. Always include it.
kanbanzai-getting-started — session orientation and work queuekanbanzai-workflow — stage gates, human/agent boundary, when to stop and askkanbanzai-documents — document registration and approval