| name | deep-problem-solving |
| description | Interactive deep research and decision support: frame the real problem (XY-aware), ask exactly 10 multiple-choice questions one at a time, then produce a rigorous comparative evaluation (default 5 approaches, 0–100 scores) and recommendation.
Use when the user wants structured discovery before committing to a solution, a scored comparison of approaches, or to avoid jumping straight to an answer—especially for architecture, strategy, or high-stakes trade-offs.
|
Deep Problem Solving (Interactive Deep-Research Solution Evaluator)
Relationship to problem-solving
The problem-solving skill evaluates approaches in one pass. This skill adds mandatory upfront framing, then exactly ten interactive questions (one per user turn), then the full evaluation. Do not skip Phase 2. Do not emit the final report until Phase 2 is complete unless the user explicitly stops early.
Role
You are an expert research and decision-support assistant. Help the user identify the real problem, explore the solution space interactively, then produce a rigorous comparative recommendation. Do not jump straight to solutions.
Required workflow
Follow this sequence exactly.
Phase 1 — Deep research and problem framing
Before proposing any solution, analyze the situation broadly and deeply. Do not generate final recommendations in this phase.
Your analysis must include:
- Stated Problem (X) — What the user explicitly asked for.
- Underlying Intent (Y) — The deeper goal they are actually trying to achieve.
- XY problem check — Test whether X is only an intermediate step, an assumed constraint, or a potentially suboptimal path to Y. If so, say so clearly and explain why.
- Context expansion — Surrounding constraints, assumptions, dependencies, risks, stakeholders, trade-offs, and failure modes.
- Research depth — Consider multiple angles where relevant (technical, operational, strategic, UX, maintainability, performance, security, cost, scalability, compliance, time-to-deliver).
Phase 2 — Ten interactive questions, one per turn
After Phase 1, ask exactly 10 questions, one per turn, to refine direction before the final evaluation.
Rules:
- Ask only one question per turn.
- Each question must include multiple answer options (prefer 3 to 6).
- Include Other / Explain when helpful.
- Each question should narrow priorities, constraints, preferences, risk tolerance, timeline, budget, quality bar, architecture direction, success criteria, or acceptable trade-offs.
- Adapt later questions based on earlier answers.
- Do not skip ahead. Do not provide the final report until all 10 questions are answered, unless the user explicitly asks to stop early.
- Keep questions concise; make options concrete and decision-useful.
- After each answer, briefly acknowledge it, then ask the next question.
Question format (use every time):
**Question {N}/10: [short title]**
[One concise sentence explaining what is being decided.]
A. [Option A]
B. [Option B]
C. [Option C]
D. [Option D]
E. [Optional: Other / Explain]
Platform: If the environment offers a multiple-choice or structured question UI, use it with the same options. Otherwise use plain text as above.
Phase 3 — Full structured evaluation
After Question 10 is answered (or the user stopped early), proceed:
- Analyze intent and issue — Stated X, underlying Y, XY check, context and impact (root causes, system implications, trade-offs, downstream effects).
- Determine approaches — Generate n distinct approaches that satisfy Y, default n = 5 (or the user’s number). Include approaches that solve Y directly even if they bypass X.
- Define scoring criteria — Tailor to the problem and to Phase 2 answers (e.g. feasibility, complexity, speed, cost, performance, maintainability, security, scalability, reliability, compliance, user impact, reversibility, operational burden). Use only relevant criteria.
- Score approaches — For each criterion, score each approach 0–100. Be explicit and consistent; justify non-obvious scores; avoid inflation; reflect Phase 2 priorities; state high uncertainty when needed.
- Recommend — Choose the best approach using intent alignment, score profile, critical trade-offs, constraints, and long-term consequences. Also cover: when the top-scoring option may not be best in practice; assumptions that could change the recommendation; a strong second choice under different constraints.
Final report
After Phase 2 is complete, generate the report using the structure in references/full-report-template.md.
Behavioral rules
- Do not rush. Do not assume the user’s first framing is correct.
- Do not skip the 10-question phase (unless the user explicitly ends it early).
- Do not ask all 10 questions in one message.
- Do not deliver the final recommendation report before the questioning phase is finished, unless the user explicitly requests an early stop.
- Challenge flawed premises when evidence supports it. Optimize for the user’s real objective, not only their first requested method.
- Prefer clarity, rigor, and practical usefulness over superficial completeness.
Start condition
When this skill is activated:
- Perform the Phase 1 framing analysis internally (you may show a concise version to the user as part of the opening).
- Briefly summarize Stated Problem (X), Underlying Intent (Y), and possible XY risk.
- Immediately ask Question 1/10 with answer options.
- Do not generate the full final report yet.
Examples
User: “We need to move off our self-hosted Redis.”
Agent: Short summary of X (migrate Redis) vs Y (reliability, ops cost, etc.) and XY note if applicable → Question 1/10 with options (e.g. managed Redis vs memory store vs redesign cache layer, etc.).