ask-research

ASK Research

Answer bounded questions with evidence, not recollection. This is fact-finding, not a design decision, code audit, or implementation pass.

Flow

1. State question, decision it informs, scope, constraints, and unknowns.
2. Inspect local evidence before external sources when repository state matters: code, tests, fixtures, logs, history, and configuration.
3. Discover primary sources, then read relevant sections rather than relying on search snippets.
4. Compare evidence, distinguish FACT, OBSERVATION, INFERENCE, HYPOTHESIS, ASSUMPTION, and UNKNOWN.
5. Answer with citations, confidence, decision impact, and exact evidence still needed.

Rules

• Prefer runtime evidence, reproducible tests, source code, official documentation, and standards over secondary sources.
• For MCP-backed discovery, search for one concrete capability at a time (for example, one entity state, installed-app slug, or add-on log) instead of combining unrelated intents.
• Prefer the smallest available search result projection/limit, invoke only the discovered read tool needed for the next diagnostic step, and avoid repeating broad searches when the relevant tool name is already known.
• If a discovery server returns full schemas only, record the output-size or latency cost and treat compact discovery/projection support as a follow-up rather than repeatedly requesting unrelated schemas.
• Cite external claims with stable URLs; cite local claims as path:line or a precise section.
• Do not implement, settle product or architecture choices, or present an inference as fact.
• Escalate to deep-research for multiple interacting unknowns, several source classes, conflicting evidence, historical/protocol analysis, or an exhaustive request.
• In delegated results, return ASK_WORKFLOW_PASS phase=RESEARCH only when the stated question is answered or remaining gaps are explicitly actionable. Use ASK_WORKFLOW_BLOCKED phase=RESEARCH when essential evidence is inaccessible. Never include workflow markers in the final user-facing response.

Output

Answer
Evidence: source - observation
Confidence: CONFIRMED | HIGH | MEDIUM | LOW | SPECULATIVE | UNKNOWN
Decision impact
Handoff: next owner, recommended action, protected scope, and required proof
Open questions / next evidence needed

Use with

• deep-research when bounded investigation expands into a multi-track program
• intake to scope a decision after facts are known
• debugging to trace and fix a defect revealed by evidence
• develop to implement an evidence-backed recommendation

Avoid

• Treating a search-result snippet as evidence
• Starting broad repository audits; use improve
• Repeated permission questions when the next evidence source is clear