Product research
One conversation, one approval, then silence until the artifact is ready. You run the front door in the user's session: the interview, the brief, the approval, the launch. Everything after the approval happens in background sessions that never ask the user anything.
Needs Claude Code with the claude command. The research runs as background sessions
(claude --bg), which a plain claude.ai chat cannot start.
This skill and every session it starts are READ ONLY on connectors. They read Slack, Drive, email, CRM and the rest; they never send, post, draft, edit, comment or react.
1. Check you can run it - silently
command -v claude git python3 && claude agents --json --all >/dev/null && echo ready
Then find night-sprint, which ships next to this skill: <this skill's base directory>/../night-sprint.
Resolve it to an absolute path - call it NS - and confirm NS/references/research-mode.md
exists.
If anything fails, stop and say exactly what is missing. Do not interview the user for a run that cannot start.
2. See which sources exist
List the information sources in your tool list, loaded or deferred: claude.ai connectors are
mcp__claude_ai_<Name>__*, other connectors are mcp__<server>__*. Web search (WebSearch,
WebFetch) is always available. Note which look relevant to the topic; you offer those in step 3.
3. Interview - short
If the user has not said what to research, ask in plain text and wait. If they have, do not ask it again.
Then ONE AskUserQuestion call, up to four questions, each with options derived from the topic:
- Decision - what will this research help decide? (e.g. whether to build it, which segment first, how to price it)
- Audience - who reads the result? (just me, the product team, leadership, a customer-facing team)
- Sources -
multiSelect. The four most relevant from step 2, with web search first. The user can name more through "Other". - Depth - Quick: 3 questions, about an hour. Standard (recommended): 5 questions, a few hours. Deep: 8 questions, overnight.
A second round only if an answer leaves the question too vague to split into sub-questions.
4. The brief - the only approval
Pick the slug: the topic in kebab case, at most 40 characters, plus -<YYYYMMDD>. The workspace
is ~/claude-research/<slug>/ - add -2, -3 if it exists. Never under ~/.claude:
Claude Code guards writes there, and a background session stops on that prompt with nobody
awake to answer it.
Write <WS>/BRIEF.md:
# <Title>
**Question:** <one sentence>
**Decision it informs:** <one sentence>
**Audience:** <who>
**Depth:** <Quick | Standard | Deep> - <n> questions, <rough duration>
## Sub-questions
1. <one line each, answerable from the allowed sources>
## Sources (read only)
- <each allowed source>
## Out of scope
- <what this will not cover>
## Deliverable
One private artifact page: the answer and what it means for the decision, findings by theme
with confidence and citations, where the evidence disagrees, gaps, and every source.
Show the brief in full in your reply. Say plainly: "After you approve, I will not ask you
anything else. The research runs in the background and you get the artifact link at the end."
Then AskUserQuestion: "Start the research with this brief?" with Approve and start and
Change something. On a change, revise the brief, show it again, and ask again.
5. Start it behind the scenes - no more questions from here
After the approval you never call AskUserQuestion again in this skill, and nothing you start
does either.
Write
<WS>/facts.envwith the keysresearch-mode.mdlists, every value on one line. Leave outTOTAL; the conductor adds it once the tickets exist.USERis the user's first name if you know it.CONTEXT_WINDOWis1000000if your own model id ends in[1m], else200000.Write
<WS>/vars-CONDUCTOR.envwith one line:NS_DIR=<NS>.Render the conductor's prompt:
bash <NS>/references/render.sh <WS> <this skill's base directory>/references/conductor-prompt.md <WS>/prompt-CONDUCTOR.txt <WS>/vars-CONDUCTOR.envLaunch it from the workspace:
cd <WS> && claude --bg -n "ns-<SLUG>-CONDUCTOR" --permission-mode auto "$(cat <WS>/prompt-CONDUCTOR.txt)"Confirm it is running: find the row named
ns-<SLUG>-CONDUCTORinclaude agents --json --alland write itssessionIdto<WS>/CONDUCTOR.session. Not into<WS>/state/: the runner treats everystate/*.sessionas a ticket.
If the launch fails - most often auto mode is not available on the account's model - do not
ask what to do. Say what failed and give one command the user can paste into a fresh terminal
to launch the conductor themselves.
6. Tell the user, then stop
Four lines, no more:
- It is running, and nothing more will be asked.
- The workspace path, and
claude agents/claude attach <short id>to look in on it. - What arrives at the end: the artifact link, in a push notification where available, in the
conductor's last message, and in
<WS>/REPORT.md. - The rough duration, from the depth.
Do not watch the run from this session. The conductor owns it.