End To End Research

End-to-End Research

Take a question from raw ask to a verified, cited deliverable. This file is the playbook — follow it in order. Load reference files on demand, not upfront:

• reference.md — the encyclopedia. Load §1 during Phase 2 (query design), §3 during Phase 4 (evidence grading), §4 during Phase 6 (contradiction resolution), §5–§9 during Phase 5 (domain-specific synthesis frameworks), §10 during Phase 7 (output templates), §11 on failures, §12 for legal/ethical boundaries.
• examples.md — worked runs (technology evaluation, company snapshot, claim verification, docs landscape, market sizing). Load when you need a concrete pattern for the current question type.
• claim-schema.json — machine contract for structured findings. Load during Phase 7 (deep tier) to validate the claims pack before handoff.

When to Use

• Any question whose answer requires multiple external sources: technology choices, vendor/company assessment, "what's the state of X", claim verification, market or competitive questions, literature surveys, due diligence.
• Trigger phrases: research, deep research, investigate, literature review, competitive analysis, market research, due diligence, fact-check, survey the field, "is it true that", "what do we know about".
• Stakes or ambiguity are high enough that a single search-and-answer would be guesswork, or the answer must be defensible — cited, dated, triangulated.

When NOT to Use

Prerequisites

• WebSearch and WebFetch tools available (free path — must work standalone).
• Memory MCP / codebase-memory available for intake checks and delivery handoffs.
• Paid APIs (exa, OpenRouter/consult, or any metered provider): require explicit user approval before first use in a session, with a cost estimate. The free path must always work alone.
• Windows host (PowerShell): shell is PowerShell. Quote URLs with single quotes ('https://...') — & and ? break bare commands. curl aliases to Invoke-WebRequest; use real tools instead. Write deliverables via the file-write tool as UTF-8, not > redirection (encoding surprises).
• Binding user rules: maximum thinking depth by default for planning, evaluation, and the verify gate; verify claims before asserting; cite sources always; no hallucinated facts — "I couldn't verify this" beats a confident guess, every time.

Operating Principles

1. Truth over speed. A late correct answer beats a fast wrong one. Never trade verification for turnaround — trade scope instead: narrow the question, keep the rigor. If evidence is thin, deliver "insufficient evidence," never plausible filler.
2. Cite everything. Every non-obvious claim in the deliverable carries an [S-###] citation into the source registry, with access date. Uncited claims get cut or relabeled as inference.
3. Triangulate. Load-bearing claims need ≥2 independent sources — different independence_group, not the same upstream origin. Ten articles quoting one press release = one source. Trace to the origin.
4. Fact ≠ claim ≠ inference. Type every finding: fact (verified against primary evidence), claim (asserted by a source, unverified), inference (your reasoning from evidence), opinion (a source's judgment), prediction (forward-looking). Never let a claim wear a fact's clothes.
5. Primary sources preferred. Trace upstream: press article → press release → filing/repo/doc/paper. Cite the most upstream source you actually fetched. Official docs, filings, source code, and datasets outrank coverage of them.
6. Recency discipline. Date-stamp every source (published_date, accessed_date) and every claim (as_of). Flag volatile claims (pricing, versions, headcount, market share) with a recheck_after horizon. For fast-moving topics, treat anything >6 months old as "verify before relying on."
7. Coverage honesty. Report what you could NOT check — failed fetches, paywalls, rate limits, skipped platforms. Compute a confidence ceiling from coverage; never present degraded coverage as comprehensive. An empty result with reasons beats padded findings.
8. No fabricated citations. Every URL in the deliverable was returned by a search or fetched THIS session. Never reconstruct a "probable" URL from memory. Model-knowledge points are labeled as such and graded E/unverified.
9. Steelman contradictions. When sources conflict, resolve per the protocol (Phase 6; full version in reference.md §4) — don't pick the convenient side.
10. Declare scope and stop conditions up front. Unbounded research is a failure mode, not diligence.

Depth Tiers

Pick a tier at intake and state it. Escalate mid-run only if findings warrant — and say so.

Defaults: fact-check → quick/standard; competitive, market, due diligence → standard/deep; literature review or "survey the field" → deep. If the user says "quick look," honor it — deliver that tier and note what a deeper tier would add.

Procedure

Intake → Plan → Collect → Evaluate → Synthesize → VERIFY GATE → Deliver → (memory handoff)
             ↑________________loop on gaps________________|

Keep the Progress Checklist (bottom of this file) updated as you move through phases. It survives context compaction and mirrors the phase gates.

Phase 1 — Intake & Scope

Produce an intake block before any search.

1. Restate the core question as one answerable sentence; if ambiguous, ask ≤3 clarifying questions — only ones that would change the plan.
2. Define success criteria: what does "answered" look like? (decision made / claim verified / landscape mapped).
3. Capture constraints: deadline, region, language, audience, budget — paid APIs (exa, OpenRouter, etc.) require explicit user approval first.
4. Pick depth tier; state it.
5. Log user-supplied assumptions as claims to verify, not facts.
6. Check memory (memory MCP / codebase-memory) for prior research on the topic — reuse settled work, re-verify stale findings.

Exit gate: one-sentence question + tier + success criteria + constraints written down.

Phase 2 — Planning

1. Decompose into 3–8 sub-questions (SQ1..SQn), each independently answerable.
2. Build a source map per sub-question: which source types can answer it (official? filings? papers? forums? code?).
3. Draft the query plan: 2–4 seed queries per sub-question, broad → narrow. Load reference.md §1 for operators, site:, filetype:, temporal filters, platform dialects.
4. Identify likely primary sources up front (vendor docs/IR page, the paper, the repo, the regulator).
5. Set stop conditions: minimum source count met AND saturation (no new material facts in last N fetches), OR time budget hit, OR success criteria met.
6. Route platforms: if the question spans surfaces (X, Reddit, HN, academic indexes), hand routing to multi-platform-agent-reach — run agent-reach doctor first if reach seems broken.

Exit gate: sub-questions + query plan + stop conditions exist. Quick tier: five lines suffice — but write them.

Phase 3 — Collection

Search → select → fetch → extract. Never cite a search snippet — fetch the page.

1. Run seed queries via WebSearch; log every query (they go in the method note).
2. Select candidates by source-type priority (primary first), not rank position.
3. WebFetch each selected source; escalate to browser/Playwright only when fetch fails (JS rendering, interaction). Respect paywalls and robots — no circumvention; use the abstract/preview or find an open equivalent.
4. Register each source (S-###) with tier, type, dates, independence_group, fetch_status.
5. Extract findings as claims (schema below): one atomic, falsifiable statement per entry, with verbatim quote + locator.
6. Chase citations upstream: when a source cites another for a load-bearing claim, fetch the origin.
7. Sharpen later queries with terminology discovered in early sources.
8. Track saturation per sub-question: 3 consecutive fetches adding nothing new → close it.
9. Check stop conditions after each sub-question, not just at the end.

Exit gate: tier's source minimum met; every sub-question has ≥1 claim or an explicit "no evidence found + why."

Phase 4 — Evaluation

Grade what you collected before building on it. Load reference.md §3 for the full evidence grading detail.

1. Assign every source a reliability grade (A–E scale below; recorded as tier T1–T5 in the claims pack).
2. Flag conflicts of interest: vendor selling the thing, competitor attacking it, sponsored content, affiliate links.
3. Verify dates: publication AND last-updated; is the claim version/time-sensitive? Mark volatile + recheck_after.
4. Assign independence_groups: sources sharing an upstream origin (same press release, same study) form ONE group — triangulation counts groups, not sources.
5. Demote or drop D/E sources unless they are the subject of study (e.g., analyzing community sentiment).

Source Reliability Scale (A–E)

Rules: key claims need A/B support or two independent C's. Never rest a conclusion on D/E. Grade the source for this claim, not the brand — a great outlet's sponsored post is D. When unsure, grade down. A source's grade (A–E ↔ T1–T5) feeds the claim's evidence grade, which also weighs independence and verification status.

Phase 5 — Synthesis

Load reference.md §5–§9 for domain frameworks: competitive (§5), market (§6), technical due diligence (§7), literature review (§8), fact-checking (§9).

1. Build the claims table: all claims grouped by sub-question, with type, grade, confidence, and as_of.
2. Build the evidence matrix for decisions: options × criteria, each cell backed by claim IDs.
3. Keep the three layers separate in the draft: facts (verified), claims (attributed, unverified), inferences (yours — written as "I infer/assess").
4. Answer each sub-question explicitly, then compose the top-level answer.
5. List open questions with reasons (no_source_found, conflicting_evidence, out_of_budget, inaccessible, too_fresh) and what evidence would settle each.
6. Surface surprises — findings contradicting the user's stated assumptions get top billing, not burial.

Phase 6 — Verification Gate

Do not skip on standard/deep. This is where hallucinations die. Load reference.md §4 for the full contradiction resolution protocol.

1. Re-check every load-bearing citation at quote level: does the fetched text actually say what the claim says?
2. Verify all numbers and dates against sources character-by-character — numbers transpose, dates drift.
3. Resolve contradictions per protocol: (1) dates — newer primary usually wins; (2) independence — is one side a citation cluster?; (3) specificity — firsthand + specific beats general; (4) unresolved → report BOTH sides with grades and dates, mark claim contested.
4. Adversarial pass (deep tier): 2–3 queries explicitly hunting disconfirming evidence ("X problems", "X criticism", "X vs").
5. Set final per-claim confidence: confirmed / probable / contested / unverified / refuted.
6. Enforce the fact rule: a finding is typed fact only if verification status is verified or spot_checked — otherwise it's a claim.
7. Kill or downgrade anything that fails — edit the synthesis; don't caveat a broken claim into staying.

Exit gate: every claim in the deliverable is cited-and-checked, or explicitly labeled inference/unverified.

Phase 7 — Deliverable

Load reference.md §10 for output templates. Load claim-schema.json if deep tier.

1. Pick the format for the audience: executive brief (quick default), full report (deep default), decision memo, annotated bibliography.
2. Structure: Answer → key findings (confidence-tagged) → evidence detail → conflicts & open questions → method note (queries run, what failed, stop condition hit) → source registry (graded, dated).
3. State the tier run and what a deeper tier would add.
4. Deep tier: emit the claims pack and validate it against claim-schema.json before any handoff.
5. Contested or high-stakes conclusions → hand off to research-report-synthesis for multi-model consensus (paid consult — ask first).
6. Offer (never force) a vault write of durable findings via the obsidian skill (durable notes folders — never a secrets folder; never a second memory product).

Claim Schema (inline)

Record findings in the claims-pack shape from Phase 3 onward. Full contract with all enums and conditional rules: claim-schema.json. Minimal valid excerpt:

{
  "meta": {
    "question": "Which managed vector DB fits a solo-dev game-companion API under $50/mo?",
    "depth_tier": "standard",
    "as_of": "2026-07-15",
    "coverage": { "source_classes_planned": 4, "source_classes_ok": 4, "degraded": [], "confidence_ceiling": "high" }
  },
  "sources": [
    {
      "id": "S-001",
      "url": "https://qdrant.tech/pricing/",
      "title": "Qdrant Cloud Pricing",
      "publisher": "Qdrant",
      "published_date": "2026-05-02",
      "accessed_date": "2026-07-15",
      "tier": "T1",
      "type": "official",
      "independence_group": "G1",
      "conflict_of_interest": "vendor pricing page — authoritative for own product, promotional elsewhere",
      "fetch_status": "fetched"
    }
  ],
  "claims": [
    {
      "id": "C-001",
      "statement": "Qdrant Cloud's free tier provides a 1GB cluster with no credit card required.",
      "sub_question": "SQ2",
      "type": "fact",
      "grade": "A",
      "confidence": "confirmed",
      "as_of": "2026-07-15",
      "volatile": true,
      "recheck_after": "2026-10-15",
      "evidence": [
        {
          "source_id": "S-001",
          "stance": "supports",
          "quote": "Start free with a 1GB cluster. No credit card required.",
          "locator": "Pricing page, free-tier card",
          "strength": "direct"
        }
      ],
      "independent_groups": 1,
      "verification": { "status": "verified", "method": "primary_source_fetch", "verified_date": "2026-07-15" }
    }
  ],
  "open_questions": [
    { "question": "Does the free tier throttle QPS?", "reason": "no_source_found", "recheck_after": "2026-08-15" }
  ]
}

Field rules (enforced by the schema): type ∈ fact | claim | inference | opinion | prediction. confidence ∈ confirmed | probable | contested | unverified | refuted. volatile: true requires recheck_after. type: "fact" requires verification status verified or spot_checked. Statements are atomic — split compound claims. A single T1 group suffices for a party's statements about its own product (as above); cross-party or load-bearing comparative claims need independent_groups ≥ 2.

Integration Points

This skill owns the pipeline and the evidence standards; those skills own routing, consensus, and persistence. Don't reimplement their internals here.

Pitfalls

• Snippet citing. Judging or quoting from search snippets without fetching. Snippets lie by truncation — fetch_status must be fetched or archived_copy for any quote.
• Citation laundering. Counting five articles that quote one press release as five sources. Independence groups exist to stop this.
• Confirmation-only search. Every query phrased to confirm the hypothesis; zero "X criticism / X failed / X vs" queries.
• Recency blindness. Presenting an old pricing page or API doc as current. Date-stamp or don't ship it.
• Confidence inflation. "X is the best" when evidence supports "X leads on criteria A and B per two B-grade sources as of July 2026."
• Inference smuggling. Your reasonable guess appearing in findings without an inference label.
• URL reconstruction. Emitting a remembered-but-unfetched URL. Grounds for failing the verify gate outright.
• Unbounded rabbit-holing. No stop conditions; the report ships late and bloated. Saturation is the signal to stop.
• Paywall/ToS circumvention. Public info only. No credential stuffing, no scraping against ToS (full boundaries: reference.md §12).
• Skipping the gate because the draft reads well. Fluency is not accuracy — plausible drafts are exactly where fabrications hide.
• Answering a different question. Scope drift from "which X for our use case" to "the history of X." Re-read the intake block before delivering.
• PowerShell URL quoting. Bare URLs with & or ? break PowerShell commands. Always use single quotes: 'https://example.com?a=1&b=2'.
• curl alias trap. On Windows, curl aliases to Invoke-WebRequest. Use WebFetch or the real tools instead; shell fetch is a last resort for binary downloads.
• File encoding. Writing deliverables via > redirection can produce unexpected encoding. Use the file-write tool as UTF-8.

Verification

After completing the pipeline, confirm:

1. Every claim is cited. Scan the deliverable: each non-obvious statement has an [S-###] tag matching a source registry entry. Uncited statements are labeled inference or cut.
2. Every URL was fetched this session. No reconstructed URLs. fetch_status is fetched or archived_copy for every quoted source.
3. Load-bearing claims triangulated. Key claims have ≥2 independent groups (different independence_group values), or a single T1 for a party's own product statements.
4. Dates present. Every source has published_date and accessed_date. Every claim has as_of. Volatile claims have recheck_after.
5. Fact rule enforced. No claim typed fact without verification.status = verified or spot_checked.
6. Contradictions resolved or reported. No silent picks — conflicts are either resolved per protocol (Phase 6) or reported as contested with both sides.
7. Coverage gaps disclosed. Method note lists failed fetches, paywalls, rate limits, skipped platforms, and the stop condition hit.
8. Deep tier: schema valid. Claims pack validates against claim-schema.json before handoff.
9. Scope intact. Re-read the intake block. The deliverable answers the one-sentence question, not a drifted variant.

Progress Checklist Template

Paste at start; keep updated (survives context compaction; mirrors the phase gates):

## Research: <one-sentence question>
Tier: quick|standard|deep · Budget: <time> · Sources: <n>/<min> · Started: <date>

- [ ] 1 Intake: question + criteria + constraints + tier
- [ ] 2 Plan: SQ1..SQn + query plan + stop conditions
- [ ] 3 Collect: SQ1 ▢ SQ2 ▢ SQ3 ▢ ... (▣ = saturated)
- [ ] 4 Evaluate: sources graded ▢ · independence groups ▢ · volatility flags ▢
- [ ] 5 Synthesize: claims table ▢ · evidence matrix ▢ · open questions ▢
- [ ] 6 Verify: quotes rechecked ▢ · numbers/dates ▢ · contradictions ▢ · adversarial pass ▢ (deep)
- [ ] 7 Deliver: format ____ · schema validated ▢ (deep) · memory handoff offered ▢
Stop condition hit: <saturation | budget | answered>
Paid APIs used: <none | list — each pre-approved>

Files

• reference.md — query design patterns (§1), source typology (§2), evidence grading A–E (§3), contradiction resolution protocol (§4), domain frameworks: competitive (§5), market (§6), technical due diligence (§7), literature review (§8), fact-checking (§9), output templates (§10), failure modes & recovery (§11), legal/ethical boundaries (§12).
• examples.md — worked runs: technology evaluation, company snapshot, claim verification, docs landscape, market sizing.
• claim-schema.json — JSON Schema for the claims pack; validate deep-tier output against it before handoff.