frank-luongt/skills/codex/deep-research

<!-- AUTO-GENERATED by export-skills.py — DO NOT EDIT -->

---

name: deep-research description: Autonomous multi-step deep research with iterative planning, gap analysis, and self-critique. Use when a topic requires exhaustive exploration beyond quick research — market intelligence, competitive landscapes, literature reviews, or strategic analyses requiring 30-60+ minutes of autonomous investigation.

Platform Note: This skill was designed for multi-agent execution. In Codex, treat sub-agent instructions as sequential steps to complete thoroughly within a single agent context.

Deep Research

Autonomous multi-step deep research inspired by Google's Gemini Deep Research architecture. Unlike quick-research (parallel breadth-first) or the research workflow (collaborative step-by-step), deep research operates as an autonomous investigator that iteratively plans, searches, reads, reasons about gaps, and self-critiques its findings before producing a comprehensive cited report.

When to Use This Skill

• Exhaustive exploration — topic requires depth, not just breadth
• Market intelligence — competitive landscapes, market sizing, trend analysis
• Literature reviews — academic, technical, or industry research
• Strategic analysis — M&A targets, investment due diligence, technology bets
• Regulatory research — compliance requirements across jurisdictions
• Technology evaluation — deep technical comparison with evidence
• Any research where "good enough" is not enough — when decisions depend on thoroughness

Do NOT Use This Skill When

• You need a quick answer (use web search directly)
• Broad but shallow exploration is sufficient (use /faos:quick-research)
• You want collaborative, guided research with human-in-the-loop (use /faos:research)
• You're exploring your own codebase (use the Explore agent)

How It Differs from Other Research Skills

| Dimension | Quick Research | Research Workflow | Deep Research | |-----------|---------------|-------------------|-------------------| | Mode | Parallel breadth-first | Collaborative step-by-step | Autonomous depth-first | | Human involvement | Minimal (scope only) | High (each step) | Minimal (scope + review) | | Duration | 5-15 min | 30-60 min | 30-90 min | | Iterations | 1-3 parallel rounds | 6 sequential steps | 3-7 adaptive iterations | | Gap analysis | Basic (spawn more if needed) | Manual (user guides) | Autonomous (self-directed) | | Self-critique | None | None | Multi-pass quality review | | Best for | Comparisons, lists, overviews | Domain/market/technical deep-dives | Exhaustive investigation | | Output quality | Good, well-cited | Comprehensive, collaborative | Authoritative, self-validated |

---

Architecture

Phase 1: PLAN          Phase 2: INVESTIGATE (iterative)        Phase 3: SYNTHESIZE
┌──────────────┐       ┌─────────────────────────────────┐     ┌──────────────────┐
│ Understand   │       │  ┌─────────┐    ┌───────────┐   │     │ Merge & Dedup    │
│ query        │──────▶│  │ Search  │───▶│ Read &    │   │────▶│ findings         │
│              │       │  │ (web +  │    │ Extract   │   │     │                  │
│ Create       │       │  │ sources)│    │ insights  │   │     │ Self-critique    │
│ research     │       │  └────┬────┘    └─────┬─────┘   │     │ (multi-pass)     │
│ plan         │       │       │               │         │     │                  │
│              │       │       ▼               ▼         │     │ Generate         │
│ Identify     │       │  ┌────────────────────────┐     │     │ final report     │
│ dimensions   │       │  │ Reason about gaps      │     │     │ with citations   │
│              │       │  │ → refine plan          │     │     │                  │
│              │       │  │ → spawn sub-agents     │     │     │ Rate confidence  │
│              │       │  │ → iterate or stop      │     │     │ per section      │
│              │       │  └────────────────────────┘     │     │                  │
└──────────────┘       └─────────── repeats 3-7x ────────┘     └──────────────────┘

Core Design Principles (Learned from Gemini Deep Research)

1. Iterative planning — the research plan evolves as findings emerge
2. Gap-driven investigation — each iteration targets what's missing, not what's known
3. Depth over breadth — follow promising threads deep before moving on
4. Self-critique before output — multiple review passes catch errors and hallucinations
5. Confidence-rated findings — every claim tagged with confidence level
6. Source triangulation — critical claims require 2+ independent sources

---

Execution

Phase 1: Plan (5-10 min)

1.1 Understand the Research Query

Analyze the user's request to extract:

• Core question — what exactly needs answering?
• Decision context — how will results be used? (investment, strategy, build vs buy, etc.)
• Constraints — geographic, temporal, industry, technology scope
• Quality bar — how authoritative do findings need to be?

If the query is ambiguous, use AskUserQuestion to clarify — but limit to 2-3 focused questions. Don't over-clarify; start researching and refine as you learn.

1.2 Create the Research Plan

Produce a structured plan with 4-8 research dimensions:

## Research Plan: [Topic]

**Core Question:** [one sentence]
**Decision Context:** [how results will be used]
**Time Budget:** [estimated duration]

### Dimensions to Investigate
1. [Dimension] — [what to find, why it matters]
2. [Dimension] — [what to find, why it matters]
3. [Dimension] — [what to find, why it matters]
...

### Source Strategy
- Primary: [official sources, research papers, government data]
- Secondary: [industry reports, expert analysis, news]
- Validation: [cross-reference strategy]

### Known Unknowns
- [What we know we don't know yet]

Show the plan to the user and proceed unless they want changes.

Phase 2: Investigate (15-60 min, 3-7 iterations)

Each iteration follows the Search → Read → Reason → Adapt loop:

2.1 Search

Use WebSearch and WebFetch to gather information. Follow these search strategies:

Start wide, then narrow:

Iteration 1: Broad queries — "[topic] overview 2026", "[topic] market size"
Iteration 2: Targeted queries — "[specific company] [topic] strategy", "[regulation] requirements"
Iteration 3+: Gap-filling — "[specific data point] source", "[conflicting claim] evidence"

Search quality rules:

• Always include current year in searches for market/industry data
• Use site-specific searches for authoritative sources: site:gov, site:edu, site:arxiv.org
• Avoid SEO content farms — prefer primary sources over aggregators
• Search in multiple languages if the topic is region-specific

2.2 Read & Extract

For each source found:

1. Fetch the full content with WebFetch
2. Extract key facts, data points, and quotes
3. Record the source URL, title, date, and credibility assessment
4. Note any claims that conflict with prior findings

2.3 Spawn Sub-Agents for Parallel Deep-Dives

When a dimension requires substantial investigation, spawn a dedicated sub-agent:

Agent tool call:
  subagent_type: "general-purpose"
  description: "Research [specific dimension]"
  prompt: |
    You are a research analyst investigating: [specific dimension]

    Context from prior research:
    [relevant findings so far]

    Your task:
    1. Find [specific information needed]
    2. Verify claims with 2+ independent sources
    3. Note any conflicting information
    4. Rate confidence: HIGH (2+ authoritative sources), MEDIUM (1 authoritative),
       LOW (indirect/inferred)

    Return structured findings with inline citations [Source Title](URL).

    Focus areas:
    - [specific sub-topic 1]
    - [specific sub-topic 2]
    - [specific sub-topic 3]

Scaling rules:

• Simple dimensions: investigate directly (no sub-agent)
• Complex dimensions: 1 sub-agent per dimension
• Very complex topics: 2-3 sub-agents per dimension with non-overlapping scope
• Maximum 5 concurrent sub-agents per iteration

2.4 Reason About Gaps

After each iteration, evaluate what's been found vs what's needed:

### Iteration [N] Gap Analysis

**Findings so far:** [summary of what we know]

**Gaps identified:**
- [ ] [Missing data point] — needed for [reason]
- [ ] [Conflicting claims] — need resolution from [source type]
- [ ] [Shallow coverage] — [dimension] needs deeper investigation

**Next iteration focus:** [what to prioritize]
**Stop criteria met?** [Yes/No — why]

2.5 Stop Criteria

Stop iterating when ANY of these conditions are met:

1. Sufficient coverage — all dimensions have HIGH or MEDIUM confidence findings
2. Diminishing returns — last iteration found <10% new information
3. Maximum iterations reached — 7 iterations is the hard limit
4. Source saturation — same sources appearing repeatedly across searches
5. Time budget exceeded — respect the estimated duration

Phase 3: Synthesize (10-20 min)

3.1 Merge and Deduplicate

1. Collect all findings from direct research and sub-agents
2. Remove duplicate information (keep the better-sourced version)
3. Organize by theme/dimension from the research plan
4. Flag any remaining contradictions

3.2 Self-Critique (Multi-Pass Review)

Run 3 critique passes before generating the final report:

Pass 1: Factual Accuracy

• Does every factual claim have a cited source?
• Are data points (numbers, dates, percentages) verified?
• Are there any claims that seem too good/bad to be true?
• Mark unverified claims with [UNVERIFIED]

Pass 2: Completeness

• Does the research answer the core question?
• Are all dimensions from the plan addressed?
• Are there obvious follow-up questions left unanswered?
• Would a domain expert find critical gaps?

Pass 3: Coherence & Bias

• Is the narrative logically structured?
• Are there hidden biases in source selection?
• Are opposing viewpoints fairly represented?
• Is the confidence rating accurate for each section?

3.3 Generate Final Report

Write the report to the configured output path. Use this structure:

# Deep Research Report: [Topic]

**Date:** [date]
**Researcher:** FAOS Deep Research Agent
**Confidence:** [Overall: HIGH/MEDIUM/LOW]
**Sources Consulted:** [N sources]
**Research Duration:** [time spent]

---

## Executive Summary

[3-5 paragraphs capturing the most important findings. Lead with the answer to the
core question. Include key data points. End with strategic implications.]

---

## Table of Contents

[Auto-generated from sections]

---

## 1. [Dimension/Theme]

[Detailed findings with inline citations]

**Confidence:** [HIGH/MEDIUM/LOW]
**Key Data Points:**
- [data point] — [source]
- [data point] — [source]

---

## 2. [Dimension/Theme]
...

---

## Key Insights & Implications

[Synthesized conclusions that go beyond individual findings.
What patterns emerge? What are the strategic implications?
What should the reader do with this information?]

---

## Limitations & Caveats

- [Known gaps in the research]
- [Areas where data quality is uncertain]
- [Biases in available sources]

---

## Recommended Follow-Up

- [ ] [Specific follow-up investigation if needed]
- [ ] [Expert consultation recommended for X]
- [ ] [Data that should be verified with primary sources]

---

## Sources

[1] [Source Title](URL) — [date accessed, credibility note]
[2] [Source Title](URL) — [date accessed, credibility note]
...

---

Configuration

Paths

• installed_path = {project-root}/.faos/custom/skills/tools/deep-research
• references_path = {installed_path}/references
• default_output_file = {output_folder}/analysis/research/deep-research-{topic}-{date}.md

Defaults

| Setting | Default | Range | |---------|---------|-------| | Max iterations | 7 | 3-10 | | Max sub-agents per iteration | 5 | 1-10 | | Max total sub-agents | 15 | 5-30 | | Self-critique passes | 3 | 1-5 | | Confidence threshold to stop | MEDIUM on all dimensions | — | | Report length | Unlimited (as needed) | — |

---

Hard Limits

• Maximum 7 research iterations (plan + 6 refinements)
• Maximum 15 total sub-agents across all iterations
• Maximum 5 concurrent sub-agents per iteration
• Stop when findings are sufficient — thoroughness, not perfection
• Token awareness — deep research uses ~30-50x more tokens than chat; budget accordingly

---

Anti-Patterns

• Boiling the ocean — don't research everything; stay focused on the core question
• Ignoring stop criteria — when sources repeat, stop searching
• Trusting single sources — critical claims need triangulation
• Over-relying on AI summaries — verify key data points from primary sources
• Skipping self-critique — the multi-pass review catches real errors
• Reporting uncertain findings as facts — always include confidence levels

---

References

• Gemini Deep Research Architecture — iterative plan-search-read-reason loop
• Anthropic Multi-Agent Research — orchestrator-worker pattern, token scaling insights
• LangChain Open Deep Research — open-source multi-step research implementation

<!-- Source: .faos/custom/skills/tools/deep-research/SKILL.md -->