UNPKG

gepa-spo

Version:

Genetic-Pareto prompt optimizer to evolve system prompts from a few rollouts with modular support and intelligent crossover

130 lines 6.09 kB
[ { "id": "polya-4step", "hint": "Apply Pólya’s 4 steps: Understand, Plan, Execute, Reflect. Restate task, outline minimal plan, do it, then brief self-check." }, { "id": "first-principles", "hint": "Reduce to primitives; identify governing constraints; rebuild solution from basics; quantify assumptions." }, { "id": "mece-structuring", "hint": "Organize content using MECE categories (mutually exclusive, collectively exhaustive); avoid overlap and cover the space." }, { "id": "crisp-dm", "hint": "For data work, follow CRISP-DM: business understanding, data understanding, prep, modeling, evaluation, deployment." }, { "id": "sixsigma-dmaic", "hint": "Process improvement: Define→Measure→Analyze→Improve→Control; baseline KPIs, verify root causes, pilot fixes, lock in control plan." }, { "id": "lean-startup-bml", "hint": "For product uncertainty, run Build–Measure–Learn loops; define MVP + falsifiable hypotheses; pivot/persevere based on evidence." }, { "id": "cynefin-sensemaking", "hint": "Classify context (obvious/complicated/complex/chaotic); choose matching approach (best practice/expert analysis/safe‑to‑fail probes/act‑sense‑respond)." }, { "id": "ooda-loop", "hint": "For fast-changing/adversarial contexts, iterate OODA: Observe, Orient, Decide, Act; shorten loop via small probes; adapt on feedback." }, { "id": "backcasting", "hint": "Start from desired end‑state; work backward to milestones, policy/tech enablers; derive near‑term actions." }, { "id": "kepner-tregoe", "hint": "Structured analysis: Situation Appraisal → Problem Analysis (Is/Is‑Not) → Decision Analysis (criteria/weights/risks) → Potential Problem Analysis." }, { "id": "decision-tree-ev", "hint": "Map decision tree; attach probabilities/payoffs; compute expected value/utility; choose max EV; consider EVPI to bound research value." }, { "id": "mcda-ahp", "hint": "When multiple criteria/stakeholders, use AHP: define hierarchy, pairwise compare, compute weights/consistency ratio, score options." }, { "id": "value-of-information", "hint": "Compute EVPI/EVSI to decide if more data is worth it; if value < cost/time, proceed without further research." }, { "id": "bayesian-updating", "hint": "Quantify belief and act under uncertainty: set prior, define likelihoods, compute posterior; act when posterior crosses decision threshold." }, { "id": "systems-cld", "hint": "Sketch causal loop diagram; identify reinforcing/balancing loops; locate leverage points; design interventions." }, { "id": "design-of-experiments", "hint": "Use DOE for learning: define factors/responses; choose design (full/fractional/orthogonal); run randomized/blocked; analyze effects." }, { "id": "monte-carlo", "hint": "Quantify uncertainty: define input distributions; run Monte Carlo; report mean, intervals; highlight tail risks." }, { "id": "sensitivity-analysis", "hint": "Test model robustness: OAT/local; global (Sobol/variance-based); report most influential inputs." }, { "id": "fermi-estimate", "hint": "When unknown quantities appear, use Fermi estimation: list assumptions, compute order‑of‑magnitude, provide range and key drivers." }, { "id": "dimensional-analysis", "hint": "For numeric/physics problems, add unit sanity checks and dimensional analysis; show conversions; flag mismatches." }, { "id": "dp-optimal-substructure", "hint": "If overlapping subproblems/optimal substructure, define recurrence, memoize/tabulate; verify correctness by induction." }, { "id": "branch-and-bound", "hint": "For combinatorial search, define bounds and branching; prune dominated regions; record incumbent; stop on proof of optimality." }, { "id": "tdd-microcycle", "hint": "Coding help uses TDD micro-cycles: specify a tiny test (in prose), implement minimally, then refactor note." }, { "id": "mape-k-control-loop", "hint": "For operational systems, set up MAPE‑K loop: Monitor→Analyze→Plan→Execute on Knowledge; define SLOs and automated actions." }, { "id": "triage-scarcity", "hint": "When resources are scarce, triage by urgency/benefit (e.g., SALT): address ‘Immediate’ first; defer lower‑benefit cases." }, { "id": "swiss-cheese", "hint": "Model layered defenses; find where ‘holes’ align; add/strengthen barriers and monitors to break accident trajectories." }, { "id": "heuristic-evaluation-nielsen10", "hint": "For UX quality, run Nielsen’s 10 heuristics; log issues with severity (0–4) and quick fixes; tackle high‑severity first." }, { "id": "causal-dag", "hint": "Draw DAG; identify confounders; apply backdoor/frontdoor criteria; specify causal estimand; avoid conditioning on colliders." }, { "id": "com-b", "hint": "Design behavior change by mapping Capability, Opportunity, Motivation → Behavior; pick matching interventions (training, prompts, incentives, environment)." }, { "id": "principled-negotiation", "hint": "Harvard method: separate people from problem, focus on interests, invent options, insist on objective criteria; know and improve BATNA." }, { "id": "proportionality-test", "hint": "Before intrusive actions: test suitability, necessity, and proportionality (narrow balancing). Document why benefits justify burdens." }, { "id": "toc-bottleneck", "hint": "Identify system constraint; exploit, subordinate, elevate; repeat; optimise only at bottleneck." } ]