Premise
Start with testable assumptions, not polished specs. Frame a problem/solution hypothesis, validate fast, and separate thinking (brainy LLMs) from codifying (dumb LLMs).
One‑Page Workflow
Frame a crisp hypothesis
User, pain, suspected cause, intervention, metric, target, timeframe, falsifier.
Example: “We believe new sellers fail to list due to fee confusion; simplifying copy will lift first‑listing completion by 20% within 14 days; invalidate if lift <10%.”
Max your human context window
Brain‑dump constraints, knowns/unknowns, non‑goals, prior attempts, risks, edge cases, success thresholds.
Rapid desk validation with Perplexity
Use Perplexity and Deep Research to map prior art, contradictions, and unknowns.
Copy the reasoning .md into a GitHub Issue to preserve claims, citations, and gaps.
Work assumptions with brainy LLMs
Surface hidden assumptions, counter‑arguments, and “cheapest falsification tests.”
Decompose into minimal, measurable experiments; call out risks and invariants.
Codify with dumb LLMs
Turn approved decisions into specs, checklists, API stubs, test scripts, and templates.
Keep prompts concrete; low temperature; feed only validated inputs.
Make assumptions first‑class in GitHub
One Issue per assumption with status (untested/validated/invalidated), evidence for/against, test owner, deadline, decision, and pasted reasoning.md.
Test small, decide fast
Concierge tests, fake doors, mocks over full builds. Timebox. Update belief based on just‑enough signal.
Promote validated bets to spec; pivot or drop invalidated ones; log learning.
Tools at a Glance
Brainy LLMs: decomposition, tradeoffs, threat modeling, experiment design, falsifiers.
Dumb LLMs: PRD skeletons, acceptance criteria, checklists, API contracts, boilerplate.
Perplexity + Deep Research: evidence trees, contradictions, market/academic scans; treat outputs as leads to verify.
Minimal Prompts
Hypothesis: “We believe [user] struggles with [problem] because [cause]. If we [solution], [metric] will improve by [target] within [timeframe]. Invalidate if [falsifier].”
Assumption tests (brainy): “List hidden assumptions. For each, propose a 1‑week, low‑cost falsification. Prioritize by risk.”
Counter‑evidence (Perplexity): “What credible evidence contradicts [claim]? Summarize sources, methods, limitations.”
Anti‑Patterns
Collecting only confirmatory evidence; no falsifiers.
Letting Deep Research become a rabbit hole; no timebox.
Jumping to code without a metric or success threshold.
Using brainy LLMs for boilerplate or dumb LLMs for strategy.
Tiny Example
Hypothesis: “Fee copy confusion blocks listings; simpler copy lifts completion by 20% in 14 days.”
Validation: Perplexity finds clarity helps but plateaus without calculators; experienced sellers less affected.
Test: A/B copy‑only vs copy+inline calculator. Outcome: +8% vs +19% → partial validation; iterate with cost preview.
Checklist
Hypothesis with metric and falsifier
Context brain‑dump logged
Deep Research reasoning.md in GitHub
Assumptions with owners and tests
Minimal experiment defined and timeboxed
Spec/checklists generated after validation
Decision and learning archived