Research Sweep

Skill: /research-sweep [topic list]

Produce a brief whose every number you would defend in public — by researching each topic independently, then paying a second agent to destroy the first one's claims.

Don't: dispatch a researcher without the preamble from research-preamble.md — an agent researching from memory looks identical to one researching from sources until you check a citation. Don't let the same agent research a topic and verify it; self-verification confirms, it does not test. Don't merge unrefuted and refuted claims into one brief without marking which is which.

Steps

Stage 1: Scope — step scope

Settle three things before dispatching anything:

1. The topic list. Each topic gets its own agent. A topic is one question with a stated angle, not a subject area — "which trigger signals empirically correlate with closed-won, and on what evidence" beats "buying signals".
2. The shared context block. Product, positioning, and audience, pasted identically into every researcher prompt so findings come back usable rather than generic.
3. The counter-position requirement. Every topic must also research the case against itself. Name it explicitly in the prompt.

Output: a topic list and a context block. Checkpoint: show Shane the topic list before spending N agents on it.

Stage 2: Research — step research — one subagent per topic

Each prompt is, in order: the preamble (with <DATE> substituted), then

You are a deep research analyst. Research one topic exhaustively and return structured findings.

# Topic: <title>
<what to cover — enumerate the specific sub-questions, named frameworks, and
originating sources to find. Be blunt about how much of this literature is
vendor content with no underlying study. Also research the counter-position.>

## <Brand> context (the brand this research serves)
<the shared context block>

Run topics in parallel. Each returns findings with sources and a per-claim confidence.

Output: N sets of findings. Checkpoint: none — flow straight into verification per topic as each lands.

Stage 3: Verify — step verify — one subagent per load-bearing claim

A claim is load-bearing if the brief's recommendation changes when it is false. Verify those; skip the rest.

Each prompt is the preamble, then:

You are an adversarial fact-checker. Your default posture is skepticism. Another
researcher, studying "<topic>", produced this claim:

CLAIM: <claim>
THEIR EVIDENCE: <evidence>
THEIR SOURCES: <urls>
THEIR CONFIDENCE: <confidence>

Try to REFUTE it. Run your own independent searches — do not just re-read their
sources. Check whether: the sources actually say what's claimed; the data is stale;
it's vendor marketing being treated as fact; it generalizes from one anecdote; a
plausible-sounding number was fabricated; the opposite is also documented somewhere
credible.

Set refuted=true if the claim is wrong, materially overstated, or unsupported by
anything you can find. When genuinely uncertain, lean toward refuted=true and
explain. Always fill correctedClaim with the version you would defend.

The uncertainty default is the point: uncertain resolves to refuted, and correctedClaim is always populated, so a survivable version exists even when the original dies.

Output: a verdict per claim. Checkpoint: report the refutation count before merging — a high rate means the research stage was too credulous, not that the brief is thin.

Stage 4: Merge — step merge

Write one brief. Refuted claims are dropped or replaced by their correctedClaim, never silently kept. State what could not be verified and what would change the recommendation.

Output: the brief.

Resume protocol

--resume-from <step> skips earlier steps. Valid steps: scope, research, verify, merge. Required state: research needs the topic list and context block; verify needs the findings; merge needs the verdicts.

Failure handling

A subagent that dies leaves its topic unresearched — say so in the brief rather than merging a silently short sweep.