メニュー
Turn research and durable things the user says into saved project records instead of chat-only answers — when to capture, where it lands, and how to report it back without pasting the document.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SOC 職業分類に基づく
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/Wolverine971/buildos-platform --skill research-captureコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Query and manage the BuildOS Supabase database. Use for debugging data issues, checking users, LLM usage, queue jobs, schema questions, or admin database operations with safety checks.
Build animations in React Native and Expo, making the decisions in the order that determines whether they feel right — should it animate, which thread it runs on, which properties, spring or timing, how the gesture hands off, how it degrades. Writes the implementation with Reanimated, Gesture Handler, Expo Router and expo-haptics. Use when animating anything in an Expo app, adding gestures, sheets, screen transitions, press feedback or haptics, or fixing motion that stutters on device. For web animation use `animate`.
Build an animation from scratch, making the decisions in the order that determines whether it feels right — should it animate at all, what purpose, which tool, which properties, which curve and duration, how it interrupts, how it exits. Writes the implementation. Use when asked to animate something, add motion, make a component feel alive, or build a transition. For critiquing existing motion use review-animations; for auditing a whole codebase use improve-animations.
| name | Research Capture |
| description | Turn research and durable things the user says into saved project records instead of chat-only answers — when to capture, where it lands, and how to report it back without pasting the document. |
| catalog_line | Save research findings and stated next steps into project records instead of losing them in chat. |
| skill_type | procedure |
| altitude | task |
| activation | progressive |
| recommended_load_format | full |
| preserve_markdown | true |
| path | apps/web/src/lib/services/agentic-chat/tools/skills/definitions/research_capture/SKILL.md |
Procedure skill at task altitude for the moment a turn produces knowledge that does not yet exist anywhere in BuildOS: findings from web research, or something durable the user stated in passing. Its job is to get that knowledge into a project record before the turn ends, and to report it back as takeaways rather than as a wall of text.
This skill is about where knowledge lands and how it is reported. It is not about document
formatting or structure — document_workspace owns that.
Load this skill when a turn is generating knowledge that would otherwise exist only in the reply.
web_search / web_visit callsDo not load this skill for a single quick lookup answered from one search, or when the user is asking about something already in their workspace. Reading existing context is not research.
The capture decision, in order. Stop at the first match.
| Condition | Call |
|---|---|
| User explicitly said not to save, or asked a throwaway question | Do not capture. Answer in chat |
| One search, one fact, answer fits in a sentence | Do not capture. Answering is enough |
| ≥2 research calls, or any comparison / landscape / pricing / options question | Capture required before the turn ends |
| User stated a future, a decision, a blocker, or a constraint not already recorded | Capture required, as the smallest record that fits (see Procedure step 4) |
| Findings extend a document that already exists | Append to it. Do not create a parallel document |
The failure this skill exists to prevent is the silent one: a turn that researches well, answers fluently, and leaves nothing behind. The user cannot see what was lost. When the call is close, capture — an unnecessary note is cheap and visible; a lost finding is neither.
search_onto_documents / list_onto_documents). Appending to the right document beats creating a
near-duplicate.## Sources section listing
every URL used. Maximum 2 new documents per turn — more than that is bloat, not thoroughness.The reply and the record do different jobs. Do not collapse them.
The document gets: the findings, the detail, the reasoning, a ## Sources section with every URL,
and an explicit unknowns section.
The reply gets:
Never paste the document body into the reply. The reply is the summary; the document is the record. If the user wants the full text they will open it.
Stop condition: the turn is not finished until the record exists. If a write fails, say so in the reply rather than answering as though it succeeded.
## Sources section when the content came from the web.document_workspace when the question is how to structure or format the document
rather than whether to write it.util.web.searchutil.web.visitonto.document.searchonto.document.listonto.document.createonto.document.updateonto.task.createweb_search calls plus two web_visit calls on vendor pricing pages. Capture bar met.create_onto_document({ project_id: "<uuid>", title: "Pricing Landscape — Competitor Research", content: "<findings, ranges, per-vendor notes>\n\n## Unknowns\n- No public pricing for two of the five vendors\n\n## Sources\n- https://…\n- https://…" })Three things stood out:
- Most competitors land between $2.5k and $6k for a comparable scope.
- Two of the five publish no pricing at all — both route to a sales call.
- The one that publishes lowest bundles hosting, so it is not a like-for-like comparison.
Detail and sources are in Pricing Landscape — Competitor Research. Still unknown: actual close prices, which nobody publishes.
update_onto_document({ document_id: "<uuid>", update_strategy: "append", content: "## Webhooks\n<findings>\n\n## Sources\n- https://…" })update_onto_task({ task_id: "<uuid>", state_key: "done" })
create_onto_task({ project_id: "<uuid>", title: "Follow up with Northwind if no reply", description: "Waiting to hear back after the intro call." })internal-default — Written 2026-07-25 from a measured failure, not from general principle. Five
identical runs of the document-from-vague-description e2e scenario split 3 pass / 2 fail: on the
failures the agent ran exactly six web_search calls, reported findings in chat, and persisted
nothing. The task-complete-cold-reference scenario fails the same way on stated futures, hitting
zero of four durable surfaces.