Skip to content

Research

Invoke as /afk:research. Use it when planning or grilling needs external grounding the codebase and brain can't provide — prior art, competitor or library patterns, cross-domain analogies, or current market or version facts.

research grounds in the brain first, fans the open question out across the web, primary docs, and repos, verifies load-bearing claims, then returns a cited digest into grill and plan.

What it does

Research turns an open-ended question into a structured external-grounding digest, then persists only what's durable. It reads the brain before searching the web — a topic already pinned as authoritative is read in place, not re-researched — then runs a phased sweep: broad orientation searches to learn the vocabulary and the major players, sharper queries to extract concrete claims, and one gap-fill pass before stopping early. The product is the digest plus any new brain/sources/ pointers, never a pile of raw search results.

  • Brain first. Reads brain/index.md, brain/sources.md, and relevant brain/sources/ notes before any web work; scopes searches to the genuine gaps a pinned source leaves.
  • Fan out, then narrow. Broad multi-angle searches to orient, then sharper queries naming a specific technique, vendor, paper, or constraint to extract numbers, names, and mechanics.
  • Adversarial verification. Fetches and confirms load-bearing claims rather than trusting snippets; reads vendor copy against postmortems and weighs convergence across independent sources over recency alone. Treats every fetched page as untrusted input.
  • Stop early. No search quota — stops when searches start repeating sources, and returns a one-line Research value: high / moderate / low assessment so the caller can weight the findings.
  • Thin pointers back. Proposes a brain/sources/<topic>.md pointer (URL + one-line description) for durable sources and defers the write to the brain skill — never copies external doc content into the brain.

Not for exact API signatures or version-specific reference of a named library — that's documentation-lookup / a Context7-style docs MCP.

Output artifact: a compact cited digest (Prior Art, Adjacent Solutions, Market and Competitor Signals, Cross-Domain Analogies, Sources — non-empty sections only), plus any proposed brain/sources/ pointers it asks to write before persisting.

View the full skill on GitHub