Deep Research Prompt SOP

A systematic process for creating deep research prompts that produce traceable, vault-contextualizable findings. Sits upstream of the Research Response SOP and downstream of the Decision Lifecycle.

Core principle: A deep research prompt serves two masters — it guides the breadth scan (Claude.ai Research) AND pre-structures the depth pass (Claude Code + vault-rag). Design for both stages, not just the first.

This connects to Question Generation SOP (where questions come from) and The Decision Lifecycle (when to research vs. when to act). See Deep Research Prompt Pipeline — Analysis for the full rationale.

Why This Matters

Without this process	With this process
All questions presented equally → research agent spreads effort thin	Must/Should/Nice tiers → agent goes deep on what matters
No decision statement → no termination condition → “is this enough?”	Decision frame → research is “enough” when the decision can be made
Generic findings → forced second-pass to contextualize everything	Traceable output format → depth pass follows URLs directly
Ad hoc depth follow-up → miss important contradictions	Pre-written second-pass brief → depth pass knows exactly what to chase
Pattern and outcome questions mixed → false certainty on unknowable things	Signal-type flags → agent says “run the experiment” instead of guessing

The Nine-Step Process

Step 1: Set Research Mode

Source: The Decision Lifecycle WEIGH stage

Classify what kind of research this is:

Mode	When to use	What it changes
Decision	A specific choice must be made	Requires Decision Frame. Has a termination condition.
Landscape	Mapping a space before decisions emerge	Uses Objective instead. No forced termination — survey is complete when the landscape is mapped.
Validation	Testing specific assumptions	Assumptions Table is the core. Questions derived from assumptions.

The test: “Am I trying to decide something, understand something, or verify something?”

Step 2: Frame the Decision or Objective

Source: The Decision Lifecycle ROUGH FRAME + Question Generation SOP Step 1

For Decision mode:

DECISION: I need to action so that result. REVERSIBILITY: Reversible / Partially reversible / Irreversible COST IF WRONG: Low / Medium / High — reason

For Landscape/Validation mode:

OBJECTIVE: I need to understand domain so that what this enables.

The test: Can you state this in one sentence? If not, you haven’t scoped it tightly enough.

Step 3: Identify Failure Modes

Source: Question Generation SOP Steps 2-3

Ask: “What could go wrong?” List 3-5 ways this decision could fail or this research could miss the mark.

Then convert each failure mode into a question: “What do I need to know to prevent this?”

Example:

Failure mode: “I choose a tool that doesn’t scale past 500 documents”
Question: “What are the scaling limits of each clustering approach at the 500-5000 document range?”

These failure-derived questions often surface the most critical research needs — things you wouldn’t think to ask from a purely positive framing.

Step 4: Add Success Criteria Questions

Source: Question Generation SOP Step 4

Beyond avoiding failure, identify what success looks like:

“What would make this choice obviously right?”
“What do practitioners recommend for my situation?”
“What does the best version of this look like?”

Step 5: Prioritize

Source: Question Generation SOP Step 5

Sort all questions into three tiers:

Tier	Criteria	Research agent behavior
MUST ANSWER	Decision cannot be made without this	Deep investigation, multiple sources
SHOULD ANSWER	Significantly improves confidence	Solid coverage, 1-2 sources
NICE TO ANSWER	Useful context, not critical	Brief treatment, deprioritize if running long

The test: “If I could only answer ONE question before acting, which would it be?” That’s your first MUST ANSWER.

Trap: If everything is MUST ANSWER, nothing is. Force yourself to differentiate.

Step 6: Flag Signal Types

Source: The Decision Lifecycle CHECK LIMIT / The Limit of Borrowed Signal

For each question, tag:

Signal Type	What it means	Research agent instruction
Pattern research	General patterns are researchable — “what do people typically do?”	Investigate thoroughly, cite sources
My-specific-outcome	Depends on my specific context — past the Limit of Borrowed Signal	Report general patterns, but flag that the answer depends on my specific setup. Recommend an experiment.

Example: “How do people handle incremental clustering?” → Pattern research. “Will incremental clustering work on MY corpus?” → My-specific-outcome.

Step 7: Extract Vault Context

Purpose: Bridge the vault-disconnect gap

The breadth agent (Claude.ai Research) can’t read your vault. Pull key specifics into the prompt:

Tool names and versions you’re already using
Architecture choices already made
Performance characteristics you’ve measured
What you’ve already tried and what happened

Keep it thin. The vault is too rich to summarize in a prompt section. This is a rough compass, not a complete map. The depth pass (Stage 2) has full vault access.

For questions with a specific baseline, add a Compare against line:

Compare against: WhisperX v0.4 on RTX 3060, currently processing 45-min episodes in ~8 min

Step 8: Write Context, Scope, and Output Format

Source: Current prompt format (these elements work well already)

Write the following sections:

Context for the Research Agent — dense situation briefing. Tech stack, what’s built, what’s known, constraints.

What I Already Have — key specifics from Step 7.

What I Do NOT Need — explicit scoping. What topics would waste effort? What’s already decided?

Output Format Requested — the traceable format:

Finding, Source, Direct URL (required), Strength of evidence
Key claims (specific, quotable — not summaries)
Comparison dimensions (what axes matter)
Open questions (what couldn’t be determined)
Contradictions

Known starting points per question — concrete entry points for the search.

Step 9: Pre-Commit Actions + Second-Pass Brief

Source: The Decision Lifecycle PREPARE

Pre-commit actions: For each assumption in the Assumptions Table, state what you’ll do if confirmed AND if refuted. For My-specific-outcome questions, specify the minimum experiment you’ll run.

Second-Pass Research Brief: Write this BEFORE the breadth scan runs. You know in advance what the disconnected agent will struggle with:

Which findings need verification against vault notes
Which URLs the depth agent should fetch and read in full
What the breadth agent likely can’t assess (your specific setup)
How to resolve contradictions between findings and existing assumptions

The Full Pipeline

The Decision Lifecycle (stages 1-3)
    ↓ feeds
Deep Research Prompt SOP (this note)
    ↓ produces
Deep Research Prompt (from template)
    ↓ run through
STAGE 1 — BREADTH: Claude.ai Research
    Scans hundreds of sources in 30-60 min
    Zero infrastructure, zero token cost beyond subscription
    ↓ produces
Findings with traceable handles
    (Direct URLs, key claims, comparison dimensions, open questions)
    ↓ guided by Second-Pass Research Brief
STAGE 2 — DEPTH: Claude Code + vault-rag + MCP servers
    Follows URLs to primary sources (WebFetch)
    Compares against vault context (vault-rag graph_search)
    Resolves open questions with full code/note access
    ↓ produces
Verified, vault-contextualized findings
    ↓ processed via
Research Response SOP
    ↓ becomes
Atomic Notes via Q-I-ST Framework

Quick Reference

The 9 Steps at a Glance:

Mode — Decision, Landscape, or Validation?
Frame — One-sentence decision or objective
Failures — What could go wrong? → Questions
Success — What does good look like? → Questions
Prioritize — Must / Should / Nice
Signal type — Pattern research or My-specific-outcome?
Vault context — Key specifics for the disconnected agent
Write it — Context, scope, output format, starting points
Pre-commit — If-then actions, second-pass brief

Time estimate: 15-30 minutes to produce a well-structured prompt.

Common Traps

Forcing decision frames on exploration. Landscape surveys don’t have a decision to frame. Use Objective mode — “I need to understand X so that Y” — not “I need to decide X.”

Everything is MUST ANSWER. If you have 8 MUST ANSWER questions, you haven’t prioritized. Force yourself: “If I could only answer three…”

Thin second-pass brief. The second-pass brief is where the depth pass gets its instructions. “Follow up on findings” is too vague. Specify which findings, which URLs, what to compare against.

Skipping “What I Don’t Need.” Without explicit scoping, the breadth agent will produce findings on adjacent topics that waste effort in both stages.

Treating breadth results as finished analysis. Claude.ai Research casts a wide net. The depth pass does the real work — following URLs, comparing against vault, resolving contradictions. Don’t skip Stage 2.

North: Where this comes from

The Decision Lifecycle (ROUGH FRAME, WEIGH, CHECK LIMIT, PREPARE)
Question Generation SOP (failure modes, success criteria, priority tiers)
Research Prompt Strategy Framework (existing Perplexity-focused methodology)

East: What opposes this

Unfocused Research (browsing without structure)
Analysis Paralysis (over-processing every research need)
Premature closure (forcing decisions before the landscape is mapped)

South: Where this leads

Better-structured prompts → higher-quality breadth scan results
Traceable output → more efficient depth passes
Pre-structured second-pass briefs → less friction, less dropped context
Verified findings → Research Response SOP → Atomic Notes

West: What’s similar

Question Generation SOP (question generation without the prompt-assembly steps)
Research Prompt Strategy Framework (10-prompt system, similar structure)
Pareto Diagnostic Framework (minimum questions, maximum insight)

Kai Mallari's Notes'

Explorer