domains / research / deep-research / deep-research

deep-research

Skill card

combined

The four copies define 'deep-research' as four distinct, non-interchangeable procedures. Imbad0202 is a 13-agent academic pipeline for systematic reviews and meta-analysis. davidondrej is a specific API wrapper for DeepAPI/Perplexity. jezweb is a pre-build product discovery workflow focused on local code exploration and competitor analysis. daymade is a multi-agent coordination system with strict asset contracts and provider lane management. No single copy is the 'original' that the others reproduce; they are different tools sharing a name. Combining them is necessary to capture the full range of 'deep research' capabilities (academic, API-based, product discovery, and coordinated multi-agent).

Repositories

Other copies

Combined skill

# Deep Research

This skill covers four distinct deep research workflows. Select the appropriate mode based on the user's intent.

## Mode 1: Academic Research (Imbad0202)

Use for systematic reviews, meta-analyses, and rigorous academic research.

1. **Routing**: Check for `[direct-mode]` token. Classify intent (explicit, cross-phase, ambiguous).
2. **Pipeline**: Dispatch to the 13-agent team for research question formulation, literature search, source verification, synthesis, and APA 7.0 report compilation.
3. **Modes**: Full research, quick brief, paper review, lit-review, fact-check, three-way literature scan, Socratic dialogue, systematic review.

## Mode 2: DeepAPI Research (davidondrej)

Use for quick, cited research reports via the DeepAPI.

1. **API Key**: Read `API_KEY` from environment. Stop if missing.
2. **Prompt**: Build a self-contained prompt with main question, context, and 3-6 sub-questions.
3. **Call**: `POST /v1/research/deep` with `Idempotency-Key` and `maxCostUsd` (default 0.70).
4. **Save**: Extract answer and sources. Save as Markdown with citations.
5. **Retry**: Handle 402, 429, and 502 errors per protocol.

## Mode 3: Product Discovery (jezweb)

Use for pre-build research on a new product or feature.

1. **Intent**: Ask what, why, constraints, and existing work.
2. **Local Exploration**: Scan user's machine for related projects, CLAUDE.md files, and reusable code.
3. **Web Research**: Search for product category, competitors, open-source alternatives, and architecture options.
4. **Ecosystem**: Research plugins, GitHub issues, forums, and app store reviews.
5. **Competitor Deep-Dive**: Analyze features, pricing, complaints, tech stack, and documentation.
6. **Brief**: Produce a research brief with recommendations.

## Mode 4: Coordinated Multi-Agent Research (daymade)

Use for high-fidelity, source-traced research reports with strict governance.

1. **Setup**: Read `research-asset-contract.md`. Create/resume a durable project study.
2. **Plan**: Draft decision questions and provider × mode lanes in `study.json`.
3. **Dispatch**: Send subagents to investigate. Keep raw search noise in task workspaces.
4. **Evidence**: Collect evidence packets with locators and excerpts. Open original sources for load-bearing claims.
5. **Synthesis**: Draft from evidence packets. Run counter-review.
6. **Verify**: Run `provider_runs.py validate` and `research_assets.py check`. Register the study.