Design & research in the AI era
Use AI to explore wider, build faster, and measure what changed. Use your judgment to decide what's worth exploring, what's worth building, and what the numbers actually mean.
The convergence trap and the amnesia problem
Ask an LLM for five options and you get five phrasings of one option. Every revision drags the output closer to the safest average. Telling it "be creative" changes nothing — creativity comes from structure, not adjectives.
Research evaporates. Teams run studies, ship decks, and start over next quarter. Last quarter's insights rot in Confluence. A PM asks "what do we know about onboarding?" and the honest answer is "let me check" — then nobody checks.
How it fits together
The double diamond (adapted from Pande's AI-era variant), four layers deep. Agents handle the breadth. You handle the decisions.
Five pieces. The Divergence Engine explores the problem space with eleven techniques that produce genuinely different AI outputs, not variations of one. Foundry picks up once you've decided what to build — agent-run, with challenge rounds. The Research Skills Pack is the playbook: method selection, screeners, synthesis, reporting. Underneath sits the Recall Model, an archive of claim capsules with evidence and confidence scores — a write path governs what gets in, a read path answers "what do we know?" before anyone starts fresh. Keystroke-Level Modeling (KLM) measures at each junction: time a task, count the clicks, turn the delta into dollars.
"Should we rebuild onboarding?"
A product team drops this question. Walk through what happens.
Before anyone diverges, ask the archive: what do we already know about onboarding? Three capsules come back — a usability study from Q1, a support ticket analysis, a competitive audit. One went stale (confidence fell below threshold). Two are current.
KLM baseline: current flow takes 47s, 23 clicksFour agents tackle the same question in isolation, each through a different lens: assumption mapper (what are we betting on?), stakeholder empathy (what does the buyer see?), question sharpener (what's actually answerable?), adversarial (why is the current flow correct?). Four genuinely different problem statements come out — not four phrasings of one.
Four versions land in front of you. The adversarial one reveals that 60% of users finish onboarding fine — the real problem is the 40% who drop at step 3. You write the brief: "Redesign step 3 to cut drop-off from 40% to under 15%."
KLM Δ cost: step 3 alone is 18s / 11 clicks — the most expensive single stepThree prototypes, built in parallel under different constraints. A: step 3 as a single ambient screen. B: step 3 as a conversation. C: eliminate step 3 entirely and infer its data from prior steps. Foundry builds them. A critic agent scores each on novelty and feasibility.
KLM score: option A = 12s, option B = 9s, option C = 0s (no step)C is fastest but loses data fidelity. B feels unfamiliar. A wins — 12 seconds, familiar pattern, lowest risk. You ship A and note that C deserves a second look once inference accuracy improves.
KLM Δ → $: 47s → 12s = 35s saved × 50K users/month × $0.004/s = $7K/monthThe decision, the rejected options, the KLM delta, and your reasoning go into a capsule. Next time someone asks about onboarding, the archive answers — with a number attached.
The pieces
Gaps
This is architecture, not practice yet. Ten versions of the recall model, zero real studies through the pipe. The divergence engine exists on paper. The skills pack is agent-ready but untested live. One pilot — one real question, end to end, failures logged honestly — closes that gap.
Beyond the pilot: who reviews capsules when the gatekeeper is out? Do topic pages hold up past a few dozen entries? Is KLM worth the overhead on low-traffic flows? What happens when the problem space is narrow enough that four lenses say the same thing?