Deep research — gather wide, lose nothing
The failure mode this skill exists to prevent: a big parallel research fan-out gathers
80–100K tokens of detail, each subagent returns a trimmed summary, the orchestrator
compresses again into a short brief, and the raw findings live only in ephemeral
agent transcripts (scratchpad task outputs) that vanish when the container is reclaimed.
Net: expensive research → a thin artifact, detail gone. Don't do that.
The rule: two durable artifacts, never just one
Every nontrivial research effort produces, and SAVES to disk:
- A synthesis — the brief / answer (tables, thesis, recommendation).
- A full appendix — the per-source / per-company / per-angle DETAIL (raw numbers,
funding histories, surfaces, dates, verbatim source URLs) the synthesis compressed.
If the saved synthesis is dramatically smaller than what was gathered, the appendix is how
you reconcile that — the detail must land somewhere durable, not only in chat or a
transcript. "I summarized it in chat" is not saved.
Running a fan-out
- Plan the angles — one subagent per distinct angle (company, source class, sub-question).
Tell the user which model each runs on and why (the subagent-model-note rule).
- Mandate write-to-disk in the subagent prompt. Every research subagent must WRITE its
full findings to a durable file (e.g. the session scratchpad dir, one file per angle)
with sources verbatim, and return only a short pointer + the headline findings. Do
NOT rely on the agent's returned summary as the record — it is trimmed by construction
and its transcript is ephemeral.
- Capture sources verbatim — every claim carries [number] [source URL] [date]
[reported vs estimate]. URLs are the first thing lost in compression; keep them.
- Synthesize from the files, not from memory — read back the written files to build
the synthesis, so nothing silently drops.
- Save both artifacts before declaring done — commit/write the synthesis AND the
appendix. Then it's saved.
Where to save
- Technical research about shipped behaviour → a
docs/ page if a user needs it to
act, otherwise a contributor note cited from CLAUDE.md keyFiles so it isn't an orphan
doc. Which tier is decided by public-vs-internal-docs and doc-tiers.
- Competitive, market, pricing, go-to-market or roadmap research → not into this
repository, in any directory. This repo is public and
no-product-strategy-here is
unconditional: a deleted file stays in history and on other branches. Hand the findings
back in the answer and let the human place them somewhere private.
- When in doubt which, ask — but never let "unsure where" become "saved nowhere."
Don't
- Don't present a one-paragraph chat summary as the deliverable for a 15-agent fan-out.
- Don't trim sources/URLs/dates to make the brief shorter — that detail goes in the appendix.
- Don't leave the only full copy in an ephemeral transcript or scratchpad you didn't commit.
1---2name: deep-research3description: Use when the user asks to research a topic in depth, map a competitive/market landscape, run a multi-source investigation, or "fan out" parallel research agents — anything where many findings must be gathered and then NOT lost. Enforces durable, detail-preserving research (write full findings to disk; keep a full appendix beside the synthesis).4---5
6# Deep research — gather wide, lose nothing
7
8The failure mode this skill exists to prevent: a big parallel research fan-out gathers
980–100K tokens of detail, each subagent returns a **trimmed summary**, the orchestrator
10**compresses again** into a short brief, and the **raw findings live only in ephemeral
11agent transcripts** (scratchpad task outputs) that vanish when the container is reclaimed.
12Net: expensive research → a thin artifact, detail gone. Don't do that.
13
14## The rule: two durable artifacts, never just one
15
16Every nontrivial research effort produces, and SAVES to disk:
17
181. **A synthesis** — the brief / answer (tables, thesis, recommendation).
192. **A full appendix** — the per-source / per-company / per-angle DETAIL (raw numbers,
20 funding histories, surfaces, dates, **verbatim source URLs**) the synthesis compressed.
21
22If the saved synthesis is dramatically smaller than what was gathered, the appendix is how
23you reconcile that — the detail must land **somewhere durable**, not only in chat or a
24transcript. "I summarized it in chat" is not saved.
25
26## Running a fan-out
27
281. **Plan the angles** — one subagent per distinct angle (company, source class, sub-question).
29 Tell the user which model each runs on and why (the subagent-model-note rule).
302. **Mandate write-to-disk in the subagent prompt.** Every research subagent must WRITE its
31 full findings to a durable file (e.g. the session scratchpad dir, one file per angle)
32 **with sources verbatim**, and return only a short pointer + the headline findings. Do
33 NOT rely on the agent's returned summary as the record — it is trimmed by construction
34 and its transcript is ephemeral.
353. **Capture sources verbatim** — every claim carries [number] [source URL] [date]
36 [reported vs estimate]. URLs are the first thing lost in compression; keep them.
374. **Synthesize from the files, not from memory** — read back the written files to build
38 the synthesis, so nothing silently drops.
395. **Save both artifacts before declaring done** — commit/write the synthesis AND the
40 appendix. Then it's saved.
41
42## Where to save
43
44- **Technical research about shipped behaviour** → a `docs/` page if a user needs it to
45 act, otherwise a contributor note cited from `CLAUDE.md` keyFiles so it isn't an orphan
46 doc. Which tier is decided by `public-vs-internal-docs` and `doc-tiers`.
47- **Competitive, market, pricing, go-to-market or roadmap research** → **not into this
48 repository, in any directory.** This repo is public and `no-product-strategy-here` is
49 unconditional: a deleted file stays in history and on other branches. Hand the findings
50 back in the answer and let the human place them somewhere private.
51- When in doubt which, ask — but never let "unsure where" become "saved nowhere."
52
53## Don't
54
55- Don't present a one-paragraph chat summary as the deliverable for a 15-agent fan-out.
56- Don't trim sources/URLs/dates to make the brief shorter — that detail goes in the appendix.
57- Don't leave the only full copy in an ephemeral transcript or scratchpad you didn't commit.
58
59<!-- vigiles:ignore-test — project-local workflow skill, not a shipped vigiles surface -->
zernie