Deep Research

Triggers

• deep research
• research this
• investigate this
• get to the bottom of
• is this true
• fact-check
• validate this claim
• what's actually going on
• separate signal from noise
• find the contrarian view
• steel-man this
• is this consensus right
• mainstream view vs reality
• dig deeper
• not satisfied with surface answer

Deep Research

Overview

Most "research" via web search returns the same answer five times because five sites copy from one source. That isn't research, it's consensus amplification. This skill exists to do the harder thing: frame the question properly, tier the evidence, steel-man the strongest dissent, and arrive at a confidence interval — not a confident-sounding paragraph.

Iron rule: never state a conclusion without (a) the source tier, (b) the strongest counter-argument and what would refute it, and (c) what you don't know.

Phase 1 — frame as a falsifiable claim

Restate the open question as one or more claims that could be wrong. Not "is X effective?" but "the claim that X reduces Y by ≥Z% in population P".

Vague questions get vague answers. The first thing to do is sharpen.

If the user's question can't be turned into a falsifiable claim, that's itself the finding — surface it ("this question is unfalsifiable as posed because…") instead of pretending to research it.

Phase 2 — set the budget

Before searching anything, write down:

• Source budget: ~6–10 sources for routine, ~20+ for high-stakes.
• Turn budget: how many web_search / web_extract calls maximum.
• Stop conditions:
  • confidence reached and stable across two independent source paths, OR
  • budget exhausted (write up the partial answer + remaining unknowns), OR
  • claim is unfalsifiable / no primary sources exist (say so).

Bounded budgets prevent runaway investigation. Budget exhaustion is a valid outcome — acknowledge what you couldn't verify rather than padding with weak sources.

Phase 3 — tiered source rubric

Score every source by tier before incorporating it into the answer.

Hard rules:

• A claim that exists only at tier 4–5 is unverified. Mark it as such. Don't escalate it just because many sites repeat it (citation laundering).
• When tier 3 cites tier 1, go read tier 1. Don't trust the summary.
• A single tier-1 source beats ten tier-4 sources saying the same thing.
• Tier-5 sources are noise unless they happen to quote a tier-1 source — in which case they're a pointer to tier 1, not a source themselves.

Use web_search (Search.sh) and web_extract to pull content; for high-stakes investigations also pull from knowledge_search(scope="learned") in case we already have prior lessons that touch this.

Phase 4 — consensus vs dissent

For each major claim, identify two camps:

1. Mainstream consensus — what most tier-2/3 sources say.
2. Strongest dissent — the most credentialed, best-evidenced voice disagreeing with consensus. Not a random contrarian. The one whose argument would actually shift consensus if it stood up.

Steel-man the dissent. Write the dissenting argument in its strongest form — what you'd say if you genuinely had to defend it. Then ask: what evidence would refute it? If that evidence exists and holds, the dissent loses. If the steel-manned dissent stands up to its own falsification test, the consensus is weaker than it looks.

This step is non-optional. If you can't find a credible dissent, that is the finding — say "no credible dissent exists in tiers 1–2, which suggests strong consensus or under-investigation by experts; flag for follow-up."

Phase 5 — cross-validation via independent paths

A claim corroborated by two sources from the same domain is barely stronger than one source. A claim corroborated by sources from independent domains (e.g., on-chain data + a court filing + a named expert) is much stronger.

For the central claim, identify ≥2 independent paths to the same conclusion. If you can only find one path, your confidence stays bounded by that single path's tier.

Phase 6 — log every step

Keep a per-investigation research log so the work is reproducible and the agent's reasoning chain is greppable. Use either:

Option A (markdown — preferred for narrative investigations): write to knowledge/learned/research/{date}-{slug}.md with frontmatter:

---
claim: "<falsifiable claim>"
verdict: supported | contradicted | inconclusive
confidence: 0.0–1.0
sources: <count>
tiers: {primary: N, expert: N, journalism: N, aggregator: N, ai: N}
dissent: <one-line steel-man of strongest dissent>
unknowns: <list of what wasn't verified>
---

(narrative: framing → sources walked → dissent → cross-validation →
verdict + confidence + caveats)

Option B (TSV — preferred for parametric / multi-claim sweeps): append to knowledge/learned/research/{slug}.tsv:

claim	source_url	tier	supports	confidence	notes

Use knowledge_write to persist; the indexer makes future knowledge_search queries surface prior investigations on the same claim. Don't re-research what you already investigated last week — check the log first.

Phase 7 — output

The deliverable is not a paragraph. It's:

1. Verdict (one sentence) — supported / contradicted / inconclusive.
2. Confidence (0.0–1.0) — what would shift it up or down.
3. Strongest evidence — the single best citation, with its tier.
4. Strongest counter — the steel-manned dissent + why it lost (or didn't).
5. What you don't know — the bounded unknowns. Always non-empty.
6. Reproduction trail — link to the research log file.

Then, if the user asked a question (vs. asking you to investigate), a 2–4 sentence direct answer that respects all the above.

Hard rules — what NOT to do

• No "I read X sources and they all agree" without naming the tiers. Five tier-4 sources agreeing is noise, not signal.
• No conclusion without a steel-manned dissent — even if you reject it, do the work of stating it.
• No undisclosed primary→summary chains. If you read a summary that cites a primary, go read the primary. If you can't access it, say so.
• No certainty leaks. If your evidence is mostly tier 3, the verdict is "supported by journalism, not yet by primary." Don't round up.
• No infinite digging. When the budget hits, stop and report, even partial. Resuming with more budget is a separate investigation.
• No pattern-matching to past cases without checking they apply. "This looks like the X situation" is a hypothesis, not a finding.

When to escalate vs continue autonomously

This skill is designed for the autonomous mind and heartbeat contexts — it does not pause to ask the user "should I keep going?". The loop runs until the budget is hit or stop conditions are met, exactly like the autonomous-experiment loop.

Escalate to the user only when:

• The claim is unfalsifiable as posed (ask for sharper question).
• All primary sources are paywalled / inaccessible (ask for access).
• The strongest dissent is credible enough that the verdict could flip with more investigation — surface it as a "taste" decision.
• Budget exhausted and confidence still <0.6 on a high-stakes claim.

Otherwise, finish the investigation, log it, return the verdict.

What this skill does NOT do

• Replace expert judgment in domains the agent isn't trained for (medical, legal, financial advice — the tier-1 source is "ask a professional," and the verdict on synthesised opinion is "do not act on this without one").
• Auto-publish findings. The output is a research log + structured verdict; acting on the verdict is a separate decision the user takes.
• Investigate live trading signals or short-horizon market movement (different skill; this one is for claims that are testable on a longer time horizon than they're tradable).

Verify

• Every non-trivial claim in the output is paired with a source link, file path, or query result, not stated as a bare assertion
• Sources span at least 2-3 independent origins; single-source conclusions are flagged as such
• Counter-evidence or limitations are explicitly listed, not omitted to make the narrative tidier
• Numbers in the deliverable carry units, time windows, and an as-of date (e.g., '$1.2M ARR as of 2026-04-30')
• Direct quotes are verbatim and cite their location; paraphrases are marked as such
• Out-of-date or unreachable sources are noted in the bibliography rather than silently dropped