| name | deep-research |
| description | Orchestrates a heavyweight fan-out research harness: decomposes a broad question into independent angles, runs parallel investigators across distinct source modalities, adversarially verifies every load-bearing claim, and synthesizes one cited, dated report. Use this skill when the user says "do deep research on X", "research this thoroughly", "fan out and research", "give me a comprehensive fact-checked report", or asks a broad or contested question that no single search thread can cover. Not for a single volatile fact, a bounded question one investigator covers end to end, or summarizing material the user supplies. |
| license | MIT |
| metadata | {"skillsmith-composes":"ground-truth-research","skillsmith-see-also":"research-note","skillsmith-invocation":"both","skillsmith-maturity":"experimental"} |
| user-invocable | true |
deep-research
A single investigator answers a question with whichever sources it
found first; a harness answers it with sources that were forced to
disagree. The cost is real — many parallel subagents — so the skill
applies when breadth, contest, or stakes justify it. A bounded
question one investigator can cover end to end is research-note's
job; a single volatile fact is an inline ground-truth-research check.
Decompose before searching
Break the question into sub-questions and pick the angles that will
answer them. Angles must be independent to be worth running in
parallel: different source modalities (official docs and specs,
source code, issue trackers and community threads, first-party
announcements and news, academic work), different stakeholders,
different time slices. Write the decomposition down before spawning
anything — it is the coverage contract. An angle dropped for cost or
time is reported as dropped, never silently absorbed.
Fan out blind
One subagent per angle, run in parallel, each blind to the others'
findings. Convergence between independent investigators is evidence;
convergence between investigators reading each other is an echo.
Each investigator returns claims, not prose: for every claim, the
source that owns it (link) and the date accessed. The
ground-truth-research discipline applies in full inside every
investigator — volatile facts verified live, every time-sensitive
claim carrying its as-of basis.
Verify adversarially
Pool the claims and dedupe them. Every load-bearing claim then gets
an independent verifier whose brief is refutation: find the primary
source and try to break the claim against it. A claim survives on
sources, not on plausibility. Claims that fail are dropped or
reported as contested with both sides cited; a claim no verifier
could check against a primary source is reported as unverified,
never quietly kept.
Ask what is missing
Before synthesis, run a completeness pass over the coverage
contract: which angle never ran, which claim is still unverified,
which cited source was never actually read. What it finds becomes
the next round of fan-out; the loop ends when a round comes back
dry, not when a round count runs out.
Synthesize one report
One report, answer first — a reader who stops after the opening
section leaves correct. Then findings with a citation per claim,
each claim labeled honestly: confirmed, single-source, contested, or
unverified. Contradictions between sources are surfaced as
contradictions, never averaged into a middle claim nobody made.
Close with what was not established and the dated source list. The
report itself carries its date — research is a snapshot. When the
user wants the result to outlive the conversation, land it as a
durable note the way research-note does and report the path.
