Files
dotfiles/pi/.pi/agent/chains/research-brief.chain.md
T
2026-07-21 21:35:16 +02:00

42 lines
1.9 KiB
Markdown

---
name: research-brief
description: Parallel web research (ollama-researcher) + local code context (scout) + synthesis. Use /run-chain research-brief "<question>" to get a single tight research brief backed by web evidence and local code.
---
# Research brief (Ollama Cloud)
A three-stage chain for "I need a well-sourced answer that combines what the web says and what our code does." Two parallel context builders, then a synthesis.
## Stages
1. **Parallel (async, fresh context)**
- `ollama-researcher``task: "Research the external side: <question>."`, `output: research/web.md`
- `scout` (built-in) — `task: "Map the local side: <question>. Identify the relevant files in this repo and how they would interact with the topic."`, `output: research/local.md`
- Concurrency 2.
2. **synthesis**`ollama-researcher` (or the parent) reads both files and writes `research/brief.md` with: direct answer, evidence from web, evidence from local code, contradictions, recommendation, and confidence level.
## When to use
- "Should we adopt library X?" — combine ecosystem evidence with our code patterns.
- "How does feature Y work in our app, and how does it compare to industry standard?"
- "What's the current best practice for Z, and where does our code diverge?"
## When NOT to use
- Pure web research (no local code angle) → just run `ollama-researcher` directly.
- Pure local question → just run `scout`.
- Implementation → use `feature-build` instead.
## Inputs
The chain task should include: the question, the repo path (for scout), and any constraints (time budget, must use free-tier models, etc.).
## Stop rules
- If the web research surfaces a fact the local code contradicts, surface it in the brief and stop before recommending.
- If the question is underspecified, parent asks the user to scope it before launching.
## Output
`research/brief.md` — direct answer first, then evidence, then recommendation, then confidence.