Live Research

Live Research

Effort: light — one background research agent reading the live source while the main lane keeps working. Removes: conclusions built on model memory that the real repo then refutes — the rework of shipping a guess.

Model memory is a guess about how a project looked at training time. The live truth is what sits on disk and in the official docs right now. This skill runs both lanes at once: while the main lane reasons about a target, a research agent reads the real thing, and its findings merge into the reasoning before any conclusion is drawn.

When to run

• You are about to reason about a project's structure, an API's contract, or a library's behavior — and you have not read the current source.
• A design, fix, or claim depends on facts that could have drifted since your training data.
• A question needs real-world facts the working context cannot answer alone.

The steps

1. Spawn the researcher in parallel. The moment reasoning about a target starts, dispatch a background research agent at the same target. The main lane keeps working; the researcher reads. Never block the work on legwork an agent can do alone.
2. Read the live truth, nearest first. The project's own README, then the section docs closest to the target, then the actual source structure — real directory listing, real file contents, real signatures. For facts outside the project: official docs, source code, specs, first-party APIs. A blog that summarizes the docs is not a primary source.
3. Stream findings back before conclusions. Findings flow to the main lane as they land, and the reasoning folds them in and corrects course. A conclusion drawn before the researcher reports on that point is a guess — mark it as one until the live truth confirms or kills it.
4. Pin every claim to the source that owns it. Each finding carries its source inline: a file path, a quoted line, a link, a commit. A claim that cannot be pinned is marked unverified, loudly — never dressed up as fact.
5. Write one cited file. Findings land in a single Markdown file, each claim with its source. Save it where the project already keeps such notes; if no convention exists, pick a sensible spot and say where, so the next agent finds it.
6. Recall before re-reading. Check notes from earlier sessions first — the same source may already be pulled. Reuse the cached finding and cite the same source. Minutes of recall beat hours of rediscovery.

Hard rules

• No conclusion before the merge. If the researcher has not reported on a point, the main lane may not state that point as settled.
• Primary sources only. Follow every claim back to the source that owns it. A secondary write-up is a pointer, not proof.
• Headless, never watched. Background research uses a headless fetch path — never a live browser a human is watching; that is a different lane.
• Unverifiable means say so. A finding with no primary source ships flagged, never silently blended into the rest.
• Zero human friction. This skill adds no approval step and no gate. It is method discipline, not a checkpoint.

What comes back

One grounded, cited Markdown file — plus a reasoning lane that was corrected mid-flight instead of after the conclusion shipped. The main lane reads the file and moves.

Works well with

• wayfinder — research tickets are the agent-alone type this skill resolves.
• root-cause-first — the same source-first discipline, aimed at bugs.
• repo-map — read the repo's map first; walk the tree raw only when the map has no answer.

Scaffold credit: Matt Pocock, research (mattpocock/skills, MIT). The composition and hard rules here are BACKS AIOS.