Deep Research

High-intensity research protocol: exhaustive discovery, recursive reflection, high-density reporting. Rigorous fact-grounding, quantitative visualization, strict structure and length.

Output

docs/dev/agents/<topic>.md — topic-specific filename, no date. Example: docs/dev/agents/auth-providers-comparison.md.

Leading word: recursion

Every search round ends with a Think (what's still unknown, what to ask next) and a Summary (what's known now). Both tight — under five lines each. The recursion drives the next round's direction.

Steps

1. Explore — minimum 10 search rounds

Each round earns its keep — new dimensions, not synonym-shuffled queries. Prefer primary sources (government, peer-reviewed, official docs). Every claim traceable.

After each round, write a short Think and Summary (see recursion above).

Parallel research: when the question splits into independent sub-questions (e.g. compare 5 auth providers, survey 3 frameworks), invoke dispatching-parallel-agents with one item per sub-question. Each subagent runs its own 10-round loop, writes to its own docs/dev/agents/<sub-topic>.md, and returns a summary. Aggregate results in step 3.

Done when: 10 rounds complete per stream; every round's Think shows diminishing new information.

2. Analyze — synthesize into markdown-native representations

• Comparison tables for entities, models, metrics
• Mermaid diagrams for flows and relationships
• Bold for key metrics and conclusions
• Data analysis via web research and codebase search

3. Persist — write to docs/dev/agents/<topic>.md

The report IS the artifact. Structure:

• Every subsection (## 3.1) must contain more than one paragraph
• Every paragraph ≥ 100 words (max 1,000)
• Tables as primary structure — aggregate recurring entities into single coherent comparison tables
• No separate "Source" column — numeric citations ([^1^]) inside data cells
• Citations: [^index^] for factual/formal pieces, max 2 per sentence
• Bold sparingly: important keywords, critical numbers, major conclusions — don't repeat the same bold entity in a short span
• References: ~10 high-quality formatted references at the end
• Conditional TL;DR at the top only if the question can be answered in a few sentences

After writing, load forge-docs — it updates docs/dev/README.md and any affected index files.

Done when: report written to output path, every claim traceable, tables/mermaid/citations in place, references formatted, and forge-docs applied.

Report modes

Style adapts to context. Default: strict academic. Omit generic Introduction/Background unless required.