research

name	research
description	Look it up before answering or building — this project's documents of record first, then external prior art. Use when the user says research, look it up, sleuth GitHub, "what do other people do", "is that a thing", "are you sure", or asks for something that has obviously been solved before.

Research

Two halves, in this order: the documents of record for this project, then prior art outside it. Answering from memory and calling it research is the thing this skill is here to stop.

Match the shape of the ask:

• A question ("is this possible", "are we doing this right", "don't touch code"): look it up, report, stop. No edits until told to go (→ audit-only).
• Part of the work ("research and fix", "sleuth GitHub and iterate until it's right"): research inside the loop. Look, apply, re-check, and say what each pass was based on — don't halt the loop to file a report.

Pick a depth and say it

Depth	For	Shape
quick	one fact, one API, one flag	a source or two, answer in a paragraph
medium	"how do people do this", choosing an approach	3-6 sources, compare, recommend
very thorough	architecture, redesign, "this has to be solved somewhere"	survey the field, name the canon, ROI-ordered options

Fan out with parallel subagents for breadth. Give each one a scope, a Thoroughness: level, Do not edit files., and a return contract: exact paths with line numbers and short quotes for code, or URL, technique, license, and applicability for outside repos.

Documents of record first

Exhaust what the project already states before searching the web.

• Tickets and specs: the tracker issue, the PRD or spec, QA feedback, the Slack or Discord thread that started it — whatever this project treats as the record.
• The repo's own planning docs: plans/, docs/, ADRs, AGENTS.md, CONTEXT.md, README.
• Figma dev mode when the question is visual. The design is the source of truth (→ ui-system).
• The repo itself: git log -S<symbol> and git log -p for why the code got this way, and the vendored source in node_modules or site-packages instead of a remembered API surface.

When two documents disagree, say so instead of silently picking one. "Double check against X, Y, Z" means re-read each one and answer per document, not re-skim your own diff.

Then prior art

Someone has almost certainly solved this. Go find them and take it.

• GitHub: gh search repos, gh search code, gh api. Read the real source, not the README. Note the license before recommending a lift.
• arXiv and papers for anything on agents, generation, simulation, scheduling, or scoring.
• Vendor docs and changelogs for API truth. Tutorials, blogs, Reddit, HN, and YouTube for what people actually hit in practice.

Report what is reusable, which pattern to copy, and what to leave behind.

Grade the evidence

Rank the sources and mark which tier a claim sits in:

1. Vendor-documented, or the shipping source code. Strongest.
2. Maintainer statements, papers, primary repos.
3. Blogs, tutorials, forum posts, model memory. Weakest, flag them as such.

Every non-obvious claim carries a URL. With no URL, write "unverified" rather than stating it flat.

Don't

• Answer from memory and call it research. "Are you sure?" and "did you research?" mean you already got caught doing it.
• Call something impossible or unsupported without opening the docs or source.
• Stop at the first hit when the question was "how do people do this".
• Hand back a link dump. The deliverable is the recommendation, with the sources under it.
• Drop the research on the floor. When it drives a change, the PR description carries what was found and why (→ pr-update). When it's a big survey with no PR yet, write it into the repo's planning docs so it outlives the chat.