Deep Research

Deep Research

Produce a decision-ready result whose material claims can be traced to inspected evidence. Derive the necessary research universe from the objective, pursue it broadly and deeply in proportion to consequence, challenge the leading conclusion, and finish the natural in-scope work without waiting for follow-up prompts.

Read for the chosen depth

• Focused: apply the contract, hierarchy, exact claim support and stopping rule below; use the relevant sections of evidence standard when evidence ambiguity or an auditable ledger matters.
• Deep or exhaustive: read coverage and research passes and evidence standard before substantive research. Map, screen, deepen, challenge and verify; source count is not coverage.
• Exhaustive coverage, exceptional consecutive maximum-reasoning passes, or joint $godmode work: also read challenger and handoff. Deep Research establishes the decision. If a separately installed $godmode skill is explicitly invoked, use it for authorized execution; otherwise complete authorized implementation and verification directly. Its absence is not a blocker.
• Material access/search failure or long, multi-surface, compaction-prone work: read the applicable recovery or checkpoint sections of access and continuity.

Inspect exact artifacts; snippets are discovery aids. Preserve protected-source exclusions on alternate routes. Challenge the strongest countercase and complete material in-scope omissions. Separate screened and deeply inspected coverage. For deep/exhaustive or consequential delivery, read research delivery; otherwise lead with the answer, exact supporting citations and material uncertainty. Instruction authority controls work, not empirical truth: test factual claims even in governing sources. Do not expand scope or force durable ledgers on small work.

Establish the Research Contract

1. Identify the decision, deliverable, or truth claim the research must support.
2. Derive scope, exclusions, time horizon, freshness, jurisdiction, acceptable uncertainty, and failure conditions from the request and governing sources.
3. Define observable acceptance criteria and what evidence could change the answer.
4. Inspect the exact named sources and the governing material applicable to the decision. Read complete operative sections and enough surrounding context to preserve meaning; expand to full documents or collections when requested or needed. Report partial coverage honestly rather than requiring an unrelated repository or account tour.
5. Make reasonable reversible assumptions when they preserve the objective; surface only assumptions that materially affect the result.

Keep research read-only unless the request separately authorizes implementation, communication, purchasing, publication, or account mutation.

Calibrate Depth

Classify the task before gathering evidence:

• Focused: resolve a bounded question with the smallest sufficient evidence.
• Deep: map the credible field, verify decisive claims, investigate contradictions, and recommend a direction.
• Exhaustive: derive the full relevant research graph, screen the broad field, deeply inspect the strongest bounded subset, run an adversarial completion pass, and create a decision-to-execution handoff in one invocation.

Use deep mode for explicit deep research. Reserve exhaustive bounded coverage for explicit exhaustive, maximum or complete research intent; keep focused, deep and exhaustive modes distinct. Do not make the user discover and request natural subquestions one at a time. Do not make a focused task exhaustive merely because more sources exist.

Apply the Evidence Hierarchy

Use current explicit instructions and governing sources to establish the requested scope and approved operating decisions. They do not make an empirical claim true. For factual support, use the strongest applicable evidence:

1. Verified live state and direct records.
2. Original authoritative evidence, including relevant governing records, first-party data, official documentation, filings, standards, source code, and primary materials.
3. Independent analysis with visible methods and incentives.
4. Credible specialist reporting.
5. Community experience and anecdotal evidence as field signal.
6. Metadata, summaries, snippets, and remembered context only as discovery aids.

Verify unstable facts live. Test incentivized claims independently when the decision warrants it. Do not inflate confidence by counting multiple retellings of one origin as independent evidence.

Stop at Decision Sufficiency

Stop when:

• acceptance criteria are met;
• all material coverage dimensions are addressed or explicitly unresolved;
• decisive claims use the strongest practical evidence;
• contradictions, inaccessible evidence, and uncertainty are visible;
• the recommendation survives the challenger pass;
• additional work is unlikely to change the decision enough to justify its cost.

Do not stop at the first plausible answer. Do not continue collecting repetitive evidence after the decision is stable.