deep-research

# 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: 1. **A synthesis** — the brief / answer (tables, thesis, recommendation). 2. **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 1. **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). 2. **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 sh...