| name | deep-research |
| type | skill |
| category | instruction |
| description | Author high-quality deep-research prompts (Gemini / ChatGPT Pro / Perplexity Deep Research), then capture the resulting documents into the PKB — including figure extraction, agent-transcribed alt-text for load-bearing images, frontmatter, and wikilink wiring to the sourcing task. |
| triggers | ["deep research","run a deep-research spike","capture deep research","pull this research into the PKB","save this gdoc"] |
| modifies_files | true |
| needs_task | true |
| mode | execution |
| domain | ["operations","research"] |
| allowed-tools | Bash,Read,Write,Edit,Glob,Grep,Skill |
| version | 1.0.0 |
| permalink | skills-deep-research |
Deep Research
Taxonomy note: This skill provides domain expertise (HOW) for the two halves of a deep-research loop: (1) authoring prompts that return usable output, and (2) capturing the resulting document into the PKB with citation fidelity and linked context. See [[aops-core/skills/remember/references/TAXONOMY.md]].
When to invoke
Two distinct entry points, often used in sequence (sometimes days apart):
- Authoring: "help me write a deep-research prompt for X", "I want to spike this question with Gemini". Route to [[prompt-authoring]].
- Capture: "I ran this deep-research, pull it in", "save this gdoc into the PKB". Route to [[pkb-capture]].
If the user gives you a URL and asks to "capture", go straight to capture — no prompt authoring needed.
The deep-research loop (overview)
- Frame the question as an affordable-loss spike task in the PKB with a well-written prompt in the body (
/planner or direct task creation).
- Run the prompt in the external tool (Gemini Deep Research, ChatGPT Pro, Perplexity Deep Research). The tool returns a Google Doc or similar.
- Capture the output back into the PKB as a
knowledge note — raw content preserved, figures transcribed, wikilinks back to the sourcing task.
- Mark the spike done with
completion_evidence pointing at the knowledge note; downstream design tasks can now consume it.
Prerequisites
rclone installed with a configured Google Drive remote named gdrive. To configure: rclone config → n (new) → name gdrive → type drive → accept defaults → authorise via browser URL. Verify with rclone lsd gdrive: (should not error).- Capture relies on
rclone for download, unzip for image extraction from .docx exports, and your own vision capability for alt-text transcription. Never route image transcription through an external service — the agent transcribes, the user verifies.
Process
Authoring a prompt
Follow [[prompt-authoring]]. Short version:
- Frame as a comparison/synthesis across 4-6 named bodies of practice (not "tell me about X")
- Require citations and flagging of thin-consensus areas
