deep-answer

name	deep-answer
description	Use when answering architecture, solution design, technical selection, or other open-ended strategy questions where direct Q&A would stay too narrow and a structured deep-research workflow is needed.

Deep Answer Skill

Overview

deep-answer is a research orchestration skill for high-level strategy prompts. It pauses before providing a quick judgment, clarifies the true decision space, and coordinates research steps so that responses expand the user’s view rather than echo the original question.

When to Use

• The user compares architectures, solution candidates, or complex technical strategies and needs trade-off visibility.
• Direct Q&A would stay narrow because multiple routes, dimensions, or benchmarks must be contrasted.
• The prompt asks for a research-style report, evidence trail, or multi-faceted recommendation rather than a single bullet answer.

When Not to Use

• The prompt is a pure fact lookup, quick debugging help, or straightforward how-to.
• Constraints are locked in and the user only needs an execution plan for a known solution.
• The user explicitly rejects structured research and wants a one-sentence verdict.

Core Principle

This skill exists to expand the user’s decision space. Every invocation should clarify boundaries, propose a structured investigation, and keep the discussion grounded in research instead of literal answers.

Workflow Stages

Stage 1: Problem Clarification

• Goal: Surface the real decision behind the user’s request, including trade-offs that matter.
• Actions: Ask follow-up questions, present 2–3 candidate interpretations, and define research boundaries, comparison dimensions, benchmarks, and success criteria.
• Deliverable: Problem Clarification Results that records the core question, study boundary, out-of-scope clarifications, dimensions to compare, relevant benchmark systems, and what success looks like.

Stage 2: Research Design

• Goal: Turn the clarified question into discrete research tasks.
• Actions: Split the investigation into 2–4 tasks (e.g., technical path, engineering cost, risk profile, organizational impact) and specify the evidence each task needs.
• Deliverable: Research Plan listing each task, its objective, expected evidence type, and preferred deliverable format.

Stage 3: Research Execution

• Goal: Gather structured evidence through agentic research or an agreed-upon alternative.
• Actions: Confirm with the user whether to launch sub-agents; if not feasible, note that the investigation will pause while preserving the plan; otherwise, execute the research tasks and capture findings.
• Deliverable: Evidence snippets and findings packets (e.g., sourced observations, data points, or quote-backed notes) that Stage 4 can synthesize.
• Constraint: Research execution is capped at two ReAct rounds. After the second round the flow must pivot to Stage 4; no additional exploratory searches are permitted.

Source Quality Gate

• Tier A: official documentation, standards, research papers, technical reports, primary data, first-party announcements.

• Tier B: high-quality engineering writeups, mature expert analysis, reputable long-form industry reporting with concrete evidence.

• Tier C: ordinary blogs, community posts, forum threads, aggregators, unsourced summaries.

• Tier D: anonymous reposts, content farms, clickbait, pure marketing copy, unverifiable opinions.

• Use rule: build key conclusions from Tier A/B sources first.

• C-source rule: Tier C sources may suggest leads, keywords, or dispute points, but may not stand alone as support for a key conclusion.

• D-source rule: Tier D sources are excluded from the evidence base.

• Verification rule: any claim that materially changes the final recommendation must be supported by either two independent high-quality sources or one strong primary source plus one independent validation source.

• Fallback rule: if evidence quality is weak, downgrade the claim to conditional or unverified and carry it into the information-gap list instead of presenting it as a settled finding.

ReAct Iteration Cap

1. Round 1 — gather initial evidence, log supporting facts, and surface the most pressing information gaps.
2. Round 2 — only revisit the highest-value gaps or verify the most contested claims before stopping exploration and proceeding to synthesis.

Stage 4: Evidence Synthesis

• Goal: Synthesize collected data into comparable conclusions.
• Actions: Deduplicate, categorize, note conflicts, and label each insight as high-confidence, conditional, or hypothesized while documenting remaining gaps.
• Evidence handling: separate core evidence from low-quality leads, record source quality when it materially affects confidence, and do not allow low-quality or single-source claims to become headline findings without explicit qualification.
• Deliverable: Research Findings summarizing key findings, route comparisons, evidence strength, major disputes, and outstanding information gaps.

Stage 5: User-Focused Answer

• Goal: Translate the research into a decision recommendation that answers the clarified question.
• Actions: Map findings to candidate routes, explain applicability conditions, justify trade-offs, and restate unknowns.
• Deliverable: Answer to the Core Question that provides the final judgement, recommended route, applicability limits, rejected options with rationale, and suggested next steps.

Stage 6: Downgrade Escalation

• Trigger: Research cannot proceed (tooling unavailable, user declines agentic execution, or evidence stalled).
• Response: Pause active investigation, preserve the current research state, and hand the user a usable downgrade package they can continue with elsewhere.
• Downgrade handoff checklist: preserve the clarification state, the current research plan, reusable prompts/tools for external work, any preliminary judgement or direction, and the most critical outstanding information gaps.

Output Contract

Every response must include the following sections with the specified fields:

1. Problem Clarification Results

• Core question
• Study boundary
• Out-of-scope items
• Comparison dimensions
• Benchmark or reference systems
• Success criteria

2. Research Findings

• Key findings
• Route comparisons
• Evidence quality
• Major disputes
• Information gaps

3. Answer to the Core Question

• Final conclusion
• Recommended route
• Applicability conditions
• Rejected route(s) with rationale
• Next steps

The final section should clearly separate what was observed (research findings) from interpretation or recommendation.

Sub-Agent Coordination

• Route Research Agent: maps candidate routes, dependencies, and scenario fit so the decision space is explicit.

• Evidence Challenge Agent: validates key claims, surfaces counterexamples, and flags boundary conditions that force deeper scrutiny.

• Engineering & Risk Agent: evaluates implementation complexity, migration cost, organizational impact, and maintenance risk so trade-offs stay grounded.

• Anti-Glue Rule: the main agent must rewrite the synthesis into a single, coherent conclusion; it may not simply paste child-agent notes into the final answer.

Anti-Pattern Guards

1. Answering only the literal question: Must expand the response with candidate routes, comparison dimensions, and benchmark context instead of stopping at the surface prompt.
2. Dumping search results without synthesis: Must turn raw evidence into structured findings before presenting them in Research Findings.
3. Stating inference as fact: Must label which statements are sourced evidence and which are the agent’s integrative judgements.
4. Searching indefinitely: Must stop after two ReAct rounds and move to synthesis, never continuing recursive follow-up queries.
5. Pasting child-agent output directly: Must rewrite sub-agent notes into a single coherent narrative; copying blocks verbatim is forbidden.
6. Giving conclusions that do not answer the user’s decision: Must tie the final section back to the clarified core question with a recommendation and trade-off rationale.
7. Treating low-quality sources as evidence: Must filter source quality before synthesis; weak or unverifiable sources may generate leads, but they cannot anchor the final recommendation.

Quality Checklist

• Did I expand routes, dimensions, and benchmark before drawing conclusions?
• Did I separate the evidence findings from the recommendation language?
• Did I filter low-quality sources out of the core evidence base and mark weak claims as conditional or unverified?
• Did I stop research after two ReAct rounds before synthesizing?
• Did the response include all three output sections with their required fields?
• If I downgraded, did I still provide a usable research plan, prompts, and gap list?