Search Drill Iterate

search-drill-iterate

A Claude Code skill for deep, iterative research. Takes a topic, an optional existing research doc, and a number of iterations. Each iteration spawns a subagent that improves the doc — without re-researching what's already covered.

How it works

An orchestrator takes three inputs — **topic**, **research file** (optional, created if absent), and **number of iterations** — fills a prompt template, and spawns one subagent per iteration.

Each subagent:

1. Reads the research doc and parses its schema sections
2. Plans which gaps to target based on Open Questions and the research questions
3. Checks Investigated Paths before every search to avoid re-research
4. Researches gaps across 10 rounds (2 cycles of 5, with a full doc re-read between)
5. Updates all schema sections before finishing

The doc is the only handoff between iterations. Each subagent starts fresh with no memory of prior iterations — it only knows what's in the file.

Document schema

Every research doc has four sections that the subagent knows how to parse and update:

Section	Purpose
Findings	The research content, organized by subtopic. Core output.
Investigated Paths	Log of what was searched and the outcome. Prevents re-research — the #1 waste of rounds.
Decision Log	Choices made during research and their rationale. Preserves reasoning across iterations.
Open Questions	Gaps to fill, with context on what was tried. UNFILLABLE: for dead ends.

See `examples/example-output.md` for a real-world example using options volatility research.

Quickstart

As a Claude Code skill:

git clone https://github.com/tiensi/search-drill-iterate.git \
  ~/.claude/skills/search-drill-iterate

Manual use with any LLM that has web search:

1. Open templates/pass.md
2. Replace placeholders: , , , ,
3. Send to the LLM
4. Run again