Every source the company already has, read in one pass.

Worked example

HYGO, a spray-planning app where 4,700 farmers plan a treatment every month. Going over the legal dose is a fineable offence, and the app said nothing: you found out at the inspection.

Sources

• Productlane

  Research repository

  Verbatims, interviews, analysis

  Worked example

  10 threads on the subject, out of 1,900 swept

  Two of them were cancellations after a failed inspection.

• PostHog

  Events & session recording

  Quantitative performance, funnels, retention…

  Worked example

  13,800 spray mixes planned in a month

  In 9% of them, someone looked the legal dose up by hand.

• PostgreSQL

  Database

  Statistics, trends

  Worked example

  One farm in four has a treatment over the legal dose

  1,600 of 6,400 active farms. 39,000 over-dose entries in 90 days.

• Claude Chrome

  Website analysis

  Structure, tone, content, UI, data

  Worked example · off

  Not read on this project.

• Notion

  Internal documentation

  Marketing, Sales, Customer Success: current and past work

  Worked example

  13 legal checks, mapped by the agronomists

  The product covered one of them.

• Google Drive

  Sheets and documents

  Shared KPIs, analysis

  Worked example · off

  Not read on this project.

• Intercom

  Customer support feedback

  Support tickets

  Worked example · off

  Not read on this project.

• NotebookLM

  Sales recording

  Raw signals from sales calls

  Worked example · off

  Not read on this project.

• Markdown files

  Product context, business rules, domain vocabulary, tone of voice

  Worked example

  The regulatory vocabulary, written down

  Buffer zones, intervals, growth stages, defined once for every agent.

• Codebase

  Access to the app codebase

  Prototypes, and iteration on production screens and flows

  Worked example

  13 alerts already live on two screens

  All of them offering a fix on a tank already emptied.

• Refero

  UI design library

  Real app and site patterns, components, flows

  Worked example · off

  Not read on this project.

• Google Meet

  Session transcripts

  Every test session, written down as it happens

  Worked example

  8 moderated sessions, transcribed live

  What people said while trying it, kept off anyone's memory.

• Maze

  Usability testing

  The prototype, run task by task, with the drop-offs counted

  Worked example

  The prototype, run task by task

  Reaching the compliance check is where people stalled.

• Figma

  Design files

  Where screens get drawn, the design system maintained, and the spec pinned

  Worked example

  The missing step, drawn

  One critique pass, five decisions, one pattern that spread.

Agents

• Claude

  Project start

  Opens the subject and finds the real problem.

  Worked example

  Opens the subject and finds the real problem.

• Claude

  Agent Researcher

  Reads every source in one pass, returns insights with their evidence.

  By hand

  • Follow-up interviews with customers
  • Session recordings, watched end to end

  Worked example

  Reads every source in one pass, returns insights with their evidence.

• Claude

  Agent Designer assistant

  Challenges flows and screens against written criteria, before a pixel is drawn.

  • Ergonomic criteria
  • Heuristics
  • Gestalt laws
  • Accessibility

  By hand

  • Iterating the screens in Figma
  • Feedback from sales, support and dev

  Worked example

  Challenges flows and screens against written criteria, before a pixel is drawn.

• Claude

  Agent Test

  Writes the protocol, reports what broke while changing it is still free.

  By hand

  • Moderating every session, live

  Worked example

  Writes the protocol, reports what broke while changing it is still free.

• Claude

  Agent PO

  Turns validated insight into scoped, prioritised decisions.

  By hand

  • Talking the scope through with the devs
  • Signing off the integration plan

  Worked example

  Turns validated insight into scoped, prioritised decisions.

Delivery

• PostHog

  Analytics setup

  Dashboard to follow and share results

  Worked example

  Measuring whether the extra step costs more saved jobs than it earns.

• Linear

  Handoff & shipping

  Writes the spec in a format a dev and their own agent can both build from.

  Worked example

  Handing over three scoped releases, each with its evidence.

What passes between them

• Problem statement

  The real problem, agreed before anyone solves it.

  • Subject
  • What is already known
  • The real question

  Worked example

  Confidence when planning a spray, and a record that holds up at inspection.

  • Subject: confidence when planning, traceability when inspected
  • Known: rivals sell this, and two customers had already left over it
  • Real question: where can a warning still change what happens?

• Insight brief

  What the sources agree on, and the proof.

  • Signals
  • Insight
  • Evidence
  • Confidence

  Worked example

  One farm in four goes over the legal dose, and finds out at the inspection.

  • Signals: 1,900 support threads read, 10 on the subject, two of them cancellations
  • Insight: the product never objects, so the inspection does it instead
  • Evidence: 1,600 of 6,400 active farms, over 90 days
  • Confidence: high. Support tickets, the database and the analytics agree.

• Design solution

  A direction, its screens, and what it costs.

  • Direction
  • Screens
  • Trade-offs

  Worked example

  Alerts while a treatment is planned, and one compliance check for the whole farm.

  • Direction: prevent, not cure. Once the tank is out, so is the wrong dose.
  • Screens: alerts on the treatment itself, and one farm-wide check to answer an inspection
  • Trade-offs: it warns, it never blocks. The job stays the farmer's to make.

• Validation report

  What broke, how badly, and what to do.

  • Protocol
  • What held
  • What broke
  • Severity
  • Recommendation

  Worked example

  8 moderated sessions, 80% of flows passed. The one that failed: the compliance check.

  • Protocol: 8 moderated sessions, the prototype run task by task
  • What held: 80% of the flows. The screens read right and the totals added up.
  • What broke: reaching the compliance check. Participants went looking elsewhere.
  • Severity: high. That is the exact screen an inspection asks to see.
  • Recommendation: put the way in where they looked, then the final spec

• Measurement plan

  How success gets measured, settled before launch.

  • Events
  • Funnel
  • Success metric

  Worked example

  Whether the warning gets acted on, and what the extra step costs.

  • Events: alert shown, correction made, treatment saved, optional field filled
  • Funnel: what share of saved treatments raised an alert, and what happened next
  • Success metric: corrections per alert, creation rate holding, optional fields filling

• Shipping spec

  What the dev team builds from.

  • Context
  • Objective
  • Scope
  • Screens
  • Acceptance

  Worked example

  Three releases, and each one waits on the data the last one captures.

  • Context: 13 legal checks apply, the product covers one
  • Objective: stop losing customers to a failed inspection
  • Release 1: ship what is already drawn, no new check
  • Release 2: ask for the crop and the growth stage, 5 checks light up
  • Release 3: the new regulatory database, 6 more