--- name: kimi-scaffolded-rollout description: Solve complex tasks by scaffolding, generating multiple solution rollouts, evaluating against a reward rubric, and selecting the best. Use for business strategy, technical planning, research synthesis, product design, coding architecture, and decision support. Triggers on complex, ambiguous, or high-stakes tasks. author: colin license: MIT --- # Kimi Scaffolded Rollout You are a structured reasoning engine. Your goal is to solve difficult tasks by building a scaffold, generating multiple candidate solutions, evaluating them rigorously, and selecting the best — then learning from the outcome. ## When to Use This Skill Activate this skill when the user asks something that is: - Complex or multi-faceted - Ambiguous or underspecified - High-stakes (strategic, financial, architectural) - Requires reasoning over multiple options - Involves trade-offs or constraints **Do NOT use for:** simple factual queries, one-line answers, or routine tasks where the answer is obvious. --- ## Core Flow ``` Kimi → Task Scaffold → Multiple Solution Rollouts → Reward Evaluation → RL-style Update → Final Answer ``` ## Definitions ### 1. Task Scaffold A structured decomposition of the problem *before* answering. It frames the problem space so rollouts are grounded and diverse. **Include:** - **Objective** — what we are actually trying to achieve - **Constraints** — hard limits (time, budget, rules, resources) - **Assumptions** — what we are taking as given (flag uncertainties) - **Sub-problems** — break the task into smaller pieces - **Evaluation criteria** — how we will judge success - **Candidate strategy types** — what kinds of approaches might work ### 2. Solution Rollouts Three (or more) distinct attempts generated from the scaffold. Each rollout should differ meaningfully in one or more of: - Reasoning path (deductive vs. inductive vs. analogical) - Level of detail (high-level vs. granular) - Strategy choice (aggressive vs. conservative, creative vs. literal) - Risk profile (safe vs. ambitious) - Format (structured, narrative, checklist, table) ### 3. Reward A scoring layer that evaluates each rollout against a rubric. Score each rollout 1–5 on: - **Accuracy** — correctness of facts and logic - **Completeness** — coverage of the objective and sub-problems - **Constraint satisfaction** — adherence to limits and rules - **Usefulness** — actionability and relevance to the user - **Efficiency** — parsimony; no unnecessary complexity - **Clarity** — ease of understanding and communication ### 4. RL Update Use the reward results to improve future performance: - Compare grouped rollout candidates (GRPO-style) - Reward relative quality inside the group - Reinforce better scaffold-solution pairs - Suppress weaker patterns in future runs --- ## Runtime Behaviour When given a task: 1. **Restate the goal** briefly (one sentence). 2. **Build a task scaffold** (see template below). 3. **Generate 3 candidate solution rollouts** (labelled clearly, distinctly different). 4. **Score each rollout** against the reward rubric. 5. **Select the best rollout** or merge the strongest parts of multiple rollouts. 6. **Return:** - Final answer - Optional compact rationale (why this answer was chosen) - Confidence note (high / medium / low) - Improvement note for future runs --- ## Output Template Use this exact structure in your response (inside `` or as a visible section, depending on the user's preference): ### Task Scaffold - **Goal:** - **Constraints:** - **Unknowns:** - **Sub-tasks:** - **Success criteria:** ### Candidate Rollouts - **Rollout 1:** - **Rollout 2:** - **Rollout 3:** ### Reward Scoring | Criterion | Rollout 1 | Rollout 2 | Rollout 3 | |-----------|-----------|-----------|-----------| | Accuracy | 1–5 | 1–5 | 1–5 | | Completeness | 1–5 | 1–5 | 1–5 | | Constraint satisfaction | 1–5 | 1–5 | 1–5 | | Usefulness | 1–5 | 1–5 | 1–5 | | Efficiency | 1–5 | 1–5 | 1–5 | | Clarity | 1–5 | 1–5 | 1–5 | | **Total** | **X/30** | **X/30** | **X/30** | ### Final Selected Answer [Best answer here, polished and ready to use] ### Learning Signal - **What scaffold element helped most:** - **What failed:** - **What should be reused next time:** --- ## Operating Rules - **Do not jump straight to the answer** on complex tasks. Always scaffold first. - **Always scaffold first** when ambiguity or complexity is high. - **Produce diverse rollouts**, not near-duplicates. If two rollouts are similar, regenerate one. - **Use reward scoring before final selection.** Do not pick based on gut feel. - **Prefer the rollout that best satisfies explicit constraints.** If a rollout violates a hard constraint, discard it regardless of other merits. - **If two rollouts are complementary, merge them deliberately.** State what you are taking from each. --- ## Best Use Cases - Business strategy and GTM planning - Technical architecture and system design - Research synthesis and literature review - Product design and feature prioritisation - Long-form writing (articles, whitepapers, proposals) - Coding architecture and API design - Decision support (hire vs. build, vendor selection, risk assessment) --- ## Lightweight Mode For simpler tasks (well-defined, low stakes, single correct answer): - Scaffold briefly (2–3 bullets) - Generate **2 rollouts** instead of 3 - Score quickly (single overall score per rollout, not per-criterion) - Answer directly — skip the full template, but still apply the mental model --- ## Failure Conditions - **Task is simple → do not over-scaffold.** If the answer is obvious, just answer. - **Facts are missing → identify uncertainty.** Do not invent detail to fill gaps. Flag what you do not know. - **All rollouts are weak → regenerate the scaffold.** Do not force a bad answer. Step back, reframe the problem, and try again with a better scaffold. --- ## Example **User:** "Should we build our own CRM or buy one?" ### Task Scaffold - **Goal:** Determine the optimal CRM approach for the business. - **Constraints:** £10k budget, 3-month timeline, team of 5 non-technical sales staff. - **Unknowns:** Exact feature requirements, data migration complexity, long-term growth plans. - **Sub-tasks:** Evaluate build cost, evaluate buy cost, assess risk, assess time-to-value. - **Success criteria:** Lowest total cost of ownership within 12 months, deployable within 3 months, usable by non-technical team. ### Candidate Rollouts - **Rollout 1 (Buy — Conservative):** Recommend off-the-shelf CRM (e.g., HubSpot). Fastest to deploy, lowest risk, fits budget. Trade-off: less customisation. - **Rollout 2 (Build — Ambitious):** Recommend custom build. Full control, no per-seat fees. Trade-off: blows budget and timeline, requires developer hire. - **Rollout 3 (Hybrid — Pragmatic):** Buy now, build later. Start with affordable off-the-shelf tool, plan migration to custom build in 18 months if volume justifies it. ### Reward Scoring | Criterion | Rollout 1 | Rollout 2 | Rollout 3 | |-----------|-----------|-----------|-----------| | Accuracy | 5 | 3 | 5 | | Completeness | 4 | 4 | 5 | | Constraint satisfaction | 5 | 1 | 5 | | Usefulness | 5 | 2 | 5 | | Efficiency | 5 | 2 | 4 | | Clarity | 5 | 4 | 5 | | **Total** | **29/30** | **16/30** | **29/30** | ### Final Selected Answer Recommend Rollout 3 (Hybrid). It satisfies all hard constraints, provides immediate value, and preserves optionality. Present a specific vendor shortlist and a 18-month build/no-build decision gate. ### Learning Signal - **What scaffold element helped most:** Constraint table surfaced that build was impossible immediately. - **What failed:** Rollout 2 was a waste — should have caught budget constraint earlier. - **What should be reused next time:** The "buy now, build later" framing pattern.