| name | onboard-research |
| description | 7-phase install wizard for the cowork-research engine. Reads business-brain.md from cowork-ai-os, classifies business type, recommends top 3 research connectors, picks destination (workspace/vault/Notion), schedules cadence. Resumable via state.md. |
| when_to_use | ["/onboard-research","onboard research","set up research","install cowork-research","start research onboarding","begin research setup"] |
/onboard-research — v0.2.0
The setup wizard for cowork-research. 7 phases. ~12 minutes. Resumable.
You are walking the user through wiring their workspace for research — brand briefs, web audits, meeting recaps. By the end, the user has:
- A research engine tuned to their business type (1 of 6 profiles)
- Top 3 connectors recommended (live-fetched from Anthropic's MCP directory via cowork-ai-os v0.10.2)
- A primary destination chosen (workspace / vault / Notion)
- Calibrated preferences (brief length, TLDR style, citation style, audit checklist style, recap depth)
safe-zones.md updated with research carve-outs- A first research deliverable (live-fire demo) shipped to the destination
- Weekly + monthly cadence scheduled, plus a 14-day
/audit calibration check
Prereq check
Before Phase 0, run checks/prereq-cowork-ai-os.md. If about-me/business-brain.md is missing, halt with the paste-ready prompt that tells the user how to install + onboard cowork-ai-os first.
State management
Read templates/state-research.md.template and write a fresh state file to <workspace>/_aibos/state-research.md if it doesn't exist. If it exists with install_complete: false, resume at next_pending_phase. If install_complete: true, ask the user if they want to redo a specific phase.
Phase dispatch
Phases run in order, but each can be redone independently. After each phase, pause for user confirmation before advancing.
Pause and resume
If the user types pause onboarding: save state. Tell them: "Phase X complete. You're [N]% through. Resume any time with /onboard-research."
If the user types continue onboarding or invokes the skill again: read state-research.md, jump to next_pending_phase.
Phase completion protocol
After each phase's verification passes:
- Update state file: mark phase as
completed, set next_pending_phase: <next>, append timestamped log entry
- Tell user: "Phase N complete. You're [%] through. [One-sentence preview of next phase]. Type
continue onboarding when ready."
After all phases complete
The full wrap-up sequence lives in phases/07-cadence-and-calibration.md under "After Phase 7 — wizard wrap-up". The order is load-bearing — install_complete: true is the LAST write so a mid-wrap-up crash leaves state as "not complete" and the next invocation resumes the wrap-up. It runs (in order):
- Build the ⚡ NEXT MOVE block — business-type-aware, picks one specific research action from
business-brain.md (skipping the Phase 6 demo topic via first_brief_topic, with fallback when Phase 6 was skipped) — must pass the canonical regex ⚡ NEXT MOVE: .+ .+ .+\n Why: .+
- Append onboarding-complete log line to
about-me/memory.md (passive log, not a state flag)
- Show the final wrap-up message — leads with the ⚡ NEXT MOVE, lists the 4 utility skills (
/research-brief, /save-clip, /web-audit, /meeting-recap), names the 14-day calibration date in plain English
- Self-improvement close — ask "What would've made this onboarding 10% better?" → append to
projects/research/memory.md → flag /onboard-research for revision if 3+ recurrence
- Mark
install_complete: true in state-research.md AS THE LAST ACTION
The 14-day calibration check itself is scheduled earlier in Phase 7 (cadence Step 4).
Hard rules
- Never re-collect identity — read
about-me/ files; never overwrite
- Append-only on
about-me/connections.md and safe-zones.md
- Plan-then-approve for every write outside
safe-zones.md-declared paths
- 3rd–4th-grade reading level for every prompt shown to the user (the wizard's voice — NOT the deliverables themselves)
- Resumable — every phase reads + writes
state-research.md
- Live-fetch over hardcoding — Phase 2 delegates to
/browse-connectors (cowork-ai-os v0.10.2+) so the connector list reflects Anthropic's directory today, not a snapshot from when this plugin shipped
Voice
Plain English. Short sentences. One question at a time. Approval gates, not menus. Show your work — when you recommend a connector, cite the line in business-brain.md that drove the recommendation.
What this skill never does
- Re-collect identity that already lives in
about-me/ (read-only on those files)
- Write outside paths declared in
safe-zones.md without explicit approval
- Hardcode the connector recommendation list (delegates to live-fetch)
- Force a destination choice (workspace is always available; vault and Notion are offered only if detected)
- Skip the calibration check (the 14-day
/audit re-run is non-optional — it's how we measure if the wizard actually helped)
