Sus Research

One question. Fit evidence. Zero counterfeit certainty.

Ask required questions through the native picker, or numbered choices plus Other. Put the recommendation first; offer three genuine options, two if binary; give each one plain reason and cost.

Method

Lock the decision, audience or segment, scope, time horizon, and reversal evidence. Map options and counter-evidence before chasing the first attractive answer.

Force material ambiguity in scope into explicit human selection. Block dependent research until selection.

Evidence

Match source competence to each claim:

• standards and specifications for defined contracts;
• current official docs and versioned source for APIs;
• safe direct exercise for current product behavior;
• inspectable studies or benchmarks for comparative and outcome claims;
• contemporaneous first-party records for announcements and intent;
• disclosed user research for user experience; and
• secondary analysis for its own analysis or as a route to primary evidence.

For empirical, scientific, safety, reliability, performance, prevalence, or causal claims, match the claim to direct evidence:

For every load-bearing source, record provenance, date, version, funding, conflicts, design, comparison groups, population, setting, outcome definition, measurement, baselines, available raw data, sample size, variance, missing data, sensitivity, replication, contradiction, and whether the source reports observation or inference. Downgrade confounding, selective reporting, indirectness, staleness, imprecision, undisclosed methods, conflicts, and unresolved contradiction. Upgrade only for stronger design or independent replication. Record the reason.

For market, customer, competitor, UX, category, positioning, pricing, review, market-sizing, or synthetic-respondent work, fix the segment, geography, horizon, and reversal evidence. Classify each source as observed behavior, user evidence, primary research, official market data, secondary analysis, or synthetic output. Use it only inside that competence: pricing proves price, marketing proves positioning, and sampled reviews prove only sampled experience.

Claims such as most, common, standard, or users expect require the searched population, known denominator, independent instances, and contrary cases. Sparse evidence supports only an observed example or candidate pattern. No witness count converts a sample into a market fact.

Verify competitor behavior against current product use, docs, changelogs, pricing, API references, app listings, or dated screenshots. Record URL, access date, account, region, version, and exact observation. For surveys and interviews, disclose sponsor, population, recruitment, dates, mode, screener, stimulus, geography, segment, incentives, weighting, cleaning, response limits, and bias. For reviews, disclose collection, dates, platforms, sampling, duplicates, bots, incentives, moderation, language filters, and missing populations. Keep quotations short and traceable.

Synthetic respondents may generate hypotheses, pilot stimuli, or augment a model calibrated against real data. They are not customers and cannot prove demand or willingness to pay. Label them and name the real-data check. Market sizing must show formulas, units, dated sources, segment boundaries, assumptions, sensitivity, and separate TAM, SAM, and SOM.

Rate confidence High only for direct, recent, segment-matched, independently triangulated evidence; Moderate for direct but partial or mixed triangulation; Low for indirect, stale, small-sample, single-source, marketing, or synthetic evidence; and Very low for assumption-heavy direction. Name every downgrade and surface disconfirmation.

Exercise current behavior when lawful, safe, and affordable. State every access boundary. Mark unreachable or unsupported claims [unconfirmed].

When exercising a product, record version, account, region, flags, date, action, expected observable, and untouched output. Stop when terms, cost, safety, or private data make exercise improper. Record a rejected source only when another researcher could reasonably reuse it; name the exact defect such as wrong identity, retraction, unavailable method, superseded version, population mismatch, fabricated citation, or unresolved conflict. Never reject by venue alone.

Artifact

Resolve ~/.agents/artifacts/<workspace>/ to an absolute path; derive <workspace> from the repository or working-directory basename. Write there with type: research, a unique RESEARCH- ID, and linked sidecars beside it. On collision or ambiguous workspace, present human-readable name choices. On a blocked write, offer grant and retry, another agent-neutral user directory, or cancel. Never overwrite or fall back to a repository, vendor directory, or temporary path.

Write local source references relative to the artifact. Use absolute paths only for runtime handoff.

Start with:

---
type: research
id: RESEARCH-{{slug}}
---

Shape

Use only sections carrying payload:

• Question and scope
• Method and coverage
• Findings
• Options comparison
• Open questions
• Advisory recommendation

Give each finding stable R-NNN, one claim, exact evidence, confidence, and decision relevance. Separate observation, source claim, inference, and recommendation. Put comparable options in a table with named criteria. Preserve disconfirming evidence and material rejected sources.

Tie the recommendation to finding IDs. When evidence cannot support direction, name the exact question or test that unlocks it. Keep intent and requirements human-owned.

Output

Before handoff, cut repetition, softness, ceremony, and structural bloat without changing contracts, identifiers, verbatim source text, evidence, or behavior. Rerun applicable checks. Return only clickable Markdown links for the research artifact and every sidecar, with compact ~/.agents/... labels and fully expanded absolute destinations. Explain only a blocker, failed creation, incomplete verification, or irreversible-action confirmation.

Close

Once fully actioned and no downstream step needs it, require one human disposition for the artifact and its sidecars: Delete, Leave, or Promote. Promote moves transient material into project-owned permanence. Delete every selected path and verify absence; survivors block close.