knowledge-base-builder
Build a project knowledge pack from Littlebird meetings and threads. Use for PRDs, decisions, and project context. Read README.md for the guide map.
Mario Aldayuz / Littlebird · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add legioncodeinc/vibe-coding-tools --skill knowledge-base-builder技能说明
name: knowledge-base-builder description: "Build a project knowledge pack from Littlebird meetings and threads. Use for PRDs, decisions, and project context. Read README.md for the guide map." license: AGPL-3.0-or-later compatibility: Claude Cowork, Claude Code 2.1 or newer, Cursor 2.4 or newer, Codex metadata: hive-tier: standalone version: "1.0.0" author: "Mario Aldayuz / Littlebird" requires: "Littlebird MCP (Power or Pro plan)"
knowledge-base-builder
Purpose
Your projects are documented in your head, in a hundred calls, and in threads nobody will read again. This skill turns that into a documentation pack that makes every future AI session productive, which is the actual payoff. A well-built knowledge base means the next agent does not need the meeting.
It ingests one project's meetings, message threads, on-screen artifacts, and documents from Littlebird capture, and writes a structured markdown pack: product requirements, architecture notes, decision records in a recognized format, a glossary, a brand brief where the project has positioning, an open questions register, and a contradiction register.
The contradiction register is the part that makes the rest trustworthy. Real projects contain conflicting statements over time: a number quoted two ways in two calls, an approach agreed and silently reversed, figures that do not match across sources. A skill that quietly picks one produces documentation that is confidently wrong, and the evidence says the model reading it will not notice. Given two passages containing contradictory facts, all models tested in the WikiContradict benchmark struggled to produce answers reflecting the conflict, and under a prompt that explicitly told them to look for contradictions the best correct rate was 43.8 percent [references/research/distilled-documentation-architecture.md section 5]. So this skill surfaces every conflict with both readings, both dates, and both receipts, and asks you to resolve it.
Mode: on-demand per project, with an optional monthly refresh routine that watches for staleness and hands back.
Littlebird MCP calls used
Real tool names, verified 2026-08-17. List the tools available in your session before calling anything; these are a starting point, not a contract.
| Tool | Used for |
|---|---|
LB_INTERNAL_LIST_MEETINGS | The project's recurring calls, by name. Also upcoming events, with a future end_date. |
LB_INTERNAL_SEARCH_MEETINGS | The topic sweep across meetings, by query |
LB_INTERNAL_GET_MEETING | The structured Decisions, Action Items, and Risks blocks, which are the spine of the pack |
LB_INTERNAL_GET_MEETING_TRANSCRIPT | Verbatim wording for a decision record quote, selectively |
search_user_context with data_source: snapshots | On-screen artifacts: designs, schemas, dashboards, docs |
search_user_context with data_source: messages | Thread-level decisions that never reached a call |
LB_INTERNAL_LIST_ROUTINES | Check for an existing refresh routine before offering to create one |
LB_INTERNAL_GET_ROUTINE_CONFIG | Read a routine's current prompt before updating it |
LB_INTERNAL_GET_ROUTINE_REPORTS | Read what the refresh routine has already flagged |
LB_INTERNAL_CREATE_ROUTINE | Create the refresh routine, after approval |
LB_INTERNAL_UPDATE_ROUTINE | Change an existing refresh routine |
LB_INTERNAL_GET_SUBSCRIPTION_STATUS | Check the plan gate and the routine limit |
A meeting lookup by NAME uses the list tool with name. A lookup by TOPIC uses the search
tool with query. Using the wrong one is the most common retrieval mistake against this
server (references/littlebird-mcp-reference.md, retrieval pattern 6).
Trigger
"Build a knowledge base for PROJECT", "document this project", "write the PRD from my calls", "what did we decide on PROJECT and when", "make a docs pack I can feed to Claude", "get this project out of my head", "write architecture notes for PROJECT", "I need to brief a contractor on PROJECT", "refresh the PROJECT knowledge base".
Not for writing a new specification for something that has not been discussed yet. This skill documents what exists in capture. With no capture there is nothing to document, and it stops.
Routine cadence
On demand per project. The pack is built in an interactive session because every material fact passes a confirmation gate and a routine cannot run gates.
Optional monthly refresh routine. It observes only: it detects that the project has moved on since the pack was built, names which documents are stale, and hands back to a Cowork session. Exact prompt text in the routine wiring section below.
Capability gate
This skill requires the Littlebird MCP on a Power or Pro plan.
- List the tools actually available in this session and use the real names you find. Do
not assume the names in
references/littlebird-mcp-reference.mdare still exact. That file is verified as of 2026-08-17 and is a starting point. - If no Littlebird tools are present, stop. Tell the user this skill needs the Littlebird MCP connected, and that it cannot build a knowledge base from a description of the project.
- If the plan gate is in doubt, call
LB_INTERNAL_GET_SUBSCRIPTION_STATUSbefore promising a pack. Check it again before offering the refresh routine, because routine count is plan-limited.
There is no degraded mode. A pack written without capture is fiction, and it is fiction that will be loaded into every future session about this project.
Do this first, every time
Read these two, in this order, before touching retrieval:
references/evidence-standards.mdfor the receipt format, the observed / inferred / external / unknown split, the attribution guardrail, and the confirmation gates.references/littlebird-mcp-reference.mdfor tool parameters, return shapes, and the limitations to design around.
Then read
阅读完整来源 ↗