Deep Research

Deep Research Skill

Core Philosophy

Most research is shallow. It covers what's easily Googlable, presents generic overviews, and stops at the surface. This skill exists to go far beyond that.

The Delivery Rule:

• Never deliver less than 7x of what was asked
• Aim for 10-15x: thorough coverage plus adjacent territory the user didn't request
• For anything in the 25x zone (interesting but outside scope), include brief pointers so nothing important stays invisible

The Quality Standard: Every section must pass the "experienced practitioner" test — would someone who's been in this space for 5 years learn something new from this output? If not, go deeper. Surface-level overviews are never acceptable as a final deliverable.

Step 1: Mode Detection

Determine the research mode before doing anything else.

Mode A: Guided (Interactive)

Trigger when: The prompt is vague, broad, ambiguous, or could go in multiple very different directions. Or the user explicitly says "ask me questions first."

Process:

1. Ask 3-5 sharp questions using the Question Decomposition technique (below)
2. Present a Research Plan and confirm
3. Execute

Question Decomposition Technique: Do not ask generic questions. Decompose the research space into decision-relevant dimensions:

• Goal Question: "What decision will this research inform?" (Not "what do you want to know" — that's circular)
• Boundary Question: "What's explicitly out of scope?" (Prevents going wide in the wrong direction)
• Knowledge Question: "What do you already know or believe about this?" (Avoids repeating known info, surfaces assumptions to test)
• Constraint Question: "Are there non-negotiable constraints?" (Budget, timeline, regulatory, technical)
• Success Question: "If this research goes perfectly, what does it let you do tomorrow?" (Reveals the real goal behind the stated goal)

Pick 3-5 of these based on what's actually ambiguous. Never ask all 5 if 3 will do.

Mode B: Autonomous (Self-Contained)

Trigger when: The prompt is specific, detailed, and provides clear context. Or the user says "just go" or gives a complete brief.

Process:

1. Internally build the Research Plan
2. Execute immediately
3. Deliver the full output

Ambiguous Cases

Default to 2 quick questions max. The user would rather get 80% right research fast than wait through 10 questions for 95% right.

Step 2: Research Planning

Build the Internal Research Plan

Every research execution builds this plan (shown to user in Guided mode, internal in Autonomous):

1. Core Question — One sentence stating exactly what we're trying to understand
2. Research Category — Classify the topic (see Playbooks in references/playbooks.md)
3. Framework Selection — Which analytical frameworks apply? (see references/frameworks.md)
4. Layer Map — What belongs in each of the 5 layers for this specific topic?
5. Source Strategy — Which sources carry the most weight? (see references/source-selection.md)
6. Known Unknowns — What do we know we don't know?
7. Hypothesized Unknown Unknowns — What might exist that hasn't been thought to look for?

Framework Selection Logic

Before executing research, select 1-3 analytical frameworks from references/frameworks.md that fit the topic. Framework selection follows this logic:

Read references/frameworks.md for full framework instructions before applying them. Do not use frameworks you haven't read the instructions for — a poorly applied framework is worse than no framework.

Step 3: Research Execution — The 5 Layers

Layer 1: Surface Scan

The easily accessible information. Table stakes.

Capture: Key players, standard definitions, basic market data, conventional wisdom, the "Wikipedia version."

Time allocation: ~10% of effort. Get this done fast and move on.

Quality gate: Can you name the top 5-10 players, define the key terms, and state the market size (if applicable) with a source? If yes, Layer 1 is done.

Layer 2: Structure & Landscape

How the space is organized — not a list, but a map with decision logic.

Capture: Market segments with real differentiators, business model variations with tradeoffs, regulatory framework, the decision tree of how options branch.

The key technique: For every option, answer three questions:

• When does this option WIN? (What conditions make it the best choice?)
• What does this option SACRIFICE? (What do you give up?)
• Who actually CHOOSES this option and why? (Real examples, not hypotheticals)

Quality gate: Could someone use this section to make a preliminary decision about which segment/approach to pursue? If not, it's too shallow.

Layer 3: Depth Dive (Where the Value Lives)

This is the most important layer. Most research stops at Layer 2. This skill keeps going.

Capture:

• Real-world case studies (successes AND failures, with specifics — names, numbers, timelines)
• Implementation gotchas and hidden costs nobody mentions upfront
• Regulatory edge cases and how they're actually enforced (vs. what's written)
• What experienced practitioners say privately vs. the official narrative
• Second-order effects ("if you do X, then Y also happens, which causes Z")
• Failure modes and the common mistakes that cause them
• The gap between marketing claims and operational reality

Where to find Layer 3 information:

• Reddit and niche forums (real practitioner experiences, search for specific terms + "reddit")
• Negative reviews on G2, Capterra, App Store (reveals real pain points)
• Government enforcement actions and audit findings
• Conference talks and podcast interviews (practitioners are more candid in conversation)
• Support forums and help documentation (reveals actual problems)
• Job postings (reveal where companies are investing and struggling)
• Regulatory comment periods (reveal industry concerns)
• Court filings and FTC actions (reveal what goes wrong)
• Academic "limitations" sections (researchers are honest about gaps)

The Absence Test: After gathering Layer 3 information, ask: "What am I NOT seeing? What should exist here but doesn't?" Absence of information is itself informative — if nobody is writing about failure cases, either the space is too new or failures are being hidden.

Quality gate: Does this section contain at least 3 specific, non-obvious insights that would surprise someone with surface-level knowledge of the topic? If not, go deeper.

Layer 4: Adjacent Opportunities (10-15x Zone)

Things the user didn't ask about but should know exist.

Identification techniques:

• Customer-outward: "Who else serves this customer, and what else do they need?"
• Solution-outward: "What other problems does this solution partially address?"
• Regulatory-outward: "What other rules/programs share the same regulatory framework?"
• Technology-outward: "What other applications use the same underlying technology?"
• Business-model-outward: "What other businesses use a similar revenue/delivery model?"

For each adjacency, provide:

• What it is (1-2 sentences)
• Why it's relevant to the original question
• Signal strength (high / medium / low)
• The one thing to search if the user wants to explore it

Quality gate: At least 3 adjacent opportunities identified. At least 1 should be genuinely surprising.

Layer 5: Horizon Pointers (25x Zone)

Brief flags for things that exist but are outside scope.

Format: 3-8 items, each with a one-liner, why it might matter, and a starting point for exploration.

Quality gate: These should be things the user would thank you for mentioning, not obvious filler.

Step 4: Synthesis

After gathering information across all 5 layers, do NOT dump findings into the output template. First, synthesize.

Read the full synthesis methodology at: references/synthesis-engine.md

The core synthesis steps are:

4a. Triangulate

Cross-reference findings from different source types. Where do 3+ independent sources agree? That's high-confidence. Where do sources contradict? That needs explicit treatment.

4b. Identify Patterns

Look across all findings for:

• Convergence patterns: Multiple unrelated sources pointing to the same conclusion
• Divergence patterns: A clear split between two camps (both may be right in different contexts)
• Silence patterns: Something that should be discussed but isn't
• Trend patterns: Directional movement across time

4c. Extract Insights (Not Just Facts)

An insight is a fact + context + implication. Transform raw findings:

• Fact: "Company X launched feature Y in 2024"
• Insight: "Company X launched feature Y in 2024, which signals that the market is moving toward Z. This matters because it changes the competitive dynamics for anyone entering now — you'd need to match Y as table stakes rather than treating it as a differentiator."

Every major finding in the output should be elevated from fact to insight.

4d. Red Team Your Own Conclusions

Before finalizing, deliberately attempt to disprove the emerging narrative. This is mandatory, not optional.

Red Team Checklist:

• What's the strongest argument AGAINST the conclusion forming in this research?
• What would need to be true for the opposite conclusion to hold?
• Is there survivorship bias? (Are we only seeing the successes?)
• Is there recency bias? (Are we overweighting recent events?)
• Is there source bias? (Are most of our sources from one perspective?)
• What would a skeptic say about this research?
• What's the most likely way this research could be WRONG?

Surface the strongest counter-argument in the output. If the research conclusion still holds after red-teaming, say so and explain why. If the red team raises genuine doubts, flag them prominently.

4e. Build the Narrative Arc

The output should tell a story, not present a filing cabinet. Before writing, identify:

• What's the single most important thing the reader should understand?
• What's the natural order of revelation? (What does the reader need to understand first before the next piece makes sense?)
• Where are the "turns" — the moments where the obvious assumption gets challenged?

Step 5: Output Construction

Format Selection

• Markdown (.md): Default for most research. Quick to read, easy to reference.
• Word document (.docx): Use when the topic is formal, the output is very long (5000+ words), or the context suggests the deliverable will be shared externally.

Document Structure

Copy code
# [Research Topic]: Deep Research Brief

## TL;DR
3-5 sentences. The most important findings. A busy person who reads only this
walks away with the key insight. This is NOT a summary of what the doc covers —
it's the answer.

## Context & Scope
What was researched, why, and what's explicitly in/out of scope. State the
frameworks used and why.

## The Landscape
Layer 1 + Layer 2 combined into a coherent narrative. Use the most appropriate
structure for the topic:
- Comparison tables for evaluating alternatives
- Decision trees for branching choices
- Market maps for competitive positioning
- Segment breakdowns for market entry
Apply the selected analytical framework(s) here.

## [Custom Title for the Depth Section]
Layer 3 content. Title this section based on the actual content:
- "Implementation Reality" for product/feature research
- "What Practitioners Actually Experience" for market research
- "Regulatory Deep Cuts" for compliance research
- "How Others Solved This" for problem-solving research

Subsections follow the natural structure of findings. Use specific case studies,
contrasting perspectives, and concrete examples.

## Adjacent Opportunities
Layer 4 content. Each item structured with:
- What it is
- Why it's relevant
- Signal strength
- Next step to explore

## Competing Perspectives & Counterarguments
Red team output. The strongest case against the emerging conclusions.
Why the research might be wrong. What would need to be true for the
opposite conclusion to hold.

## Decision Points
(If the research informs a decision)
Crystallize key tradeoffs into clearly framed choices:
- Option A: [description] → leads to [outcome], sacrifices [tradeoff]
- Option B: [description] → leads to [outcome], sacrifices [tradeoff]
Not recommendations unless explicitly asked — just cleanly framed decisions
with consequences.

## Horizon: Worth Knowing About
Layer 5 content. Brief pointers to the 25x zone.

## Sources & Confidence
What sources were used, why chosen, confidence level on key claims.
Flag anything where information was thin, contradictory, or limited.
Rate overall confidence: HIGH / MEDIUM / LOW with reasoning.

Writing Standards

• Write in prose, not bullet dumps. Bullets only for genuinely list-like content
• Every substantive claim gets a source or confidence level
• Competing perspectives sit side by side, not buried
• Numbers always get context (market size needs growth rate, comparable benchmarks, or "so what")
• Tables for comparisons (2+ options, 3+ dimensions)
• Section lengths proportional to importance, not equal
• No filler phrases ("in today's rapidly evolving landscape," "it's important to note")
• No jargon without definition on first use
• Active voice, direct statements
• When uncertain, say so explicitly with reasoning

Execution Checklist (Verify Before Delivering)

• TL;DR could stand alone and answers the actual question
• At least one analytical framework was applied (not just mentioned)
• Layer 3 exists and contains 3+ non-obvious insights
• At least 3 adjacent opportunities surfaced (Layer 4)
• Horizon pointers included (Layer 5)
• Red team step completed — counter-arguments are in the document
• Competing perspectives represented
• Sources and confidence levels noted throughout
• Synthesis happened — output contains insights, not just facts
• Document tells a narrative, not a data dump
• Someone with zero knowledge could follow the document
• Someone with deep knowledge would learn something new
• Output is genuinely 10x+ of a surface-level answer
• No section is generic filler
• The "so what" is clear for every major finding

Reference Files

Read these BEFORE executing research. They contain detailed methodology that dramatically improves output quality.

What NOT to Do

• Never deliver a generic overview and call it research
• Never present information without context or "so what"
• Never skip Layer 3 — this is where the value lives
• Never skip the red team step — unchallenged conclusions are unreliable
• Never present a single narrative without acknowledging competing views
• Never use filler phrases or hedge unnecessarily
• Never dump bullets without synthesis
• Never present data without sourcing or confidence framing
• Never stop at what was asked — always deliver the 10-15x
• Never apply a framework superficially — a checkbox PESTEL is worse than no PESTEL
• Never assume the first narrative that emerges is correct