Steer · Chapter 01

Discovery

Finding what to build: hypothesis, test, counterfactual, decision.

The first decision in automation is which work deserves to become easier. Think of a recent day when the business served a customer well. What happened for them, what did it take from the team, and where did the work get stuck? Those details give a proposed system a purpose.

There are two uncertainties to resolve: whether a process is worth improving, and whether someone values its output. A busy process may serve an unwanted offer. A promising offer may depend on work that is already cheap and reliable. Look for evidence on both sides before investing in machinery.

Choose an outcome you can observe

List a few recurring jobs and observe actual runs. Count frequency, elapsed time, review, rework and the consequences of an error. Google’s SRE work on eliminating toil offers a useful discipline here: repetitive work is a candidate for improvement, but the benefit still has to justify implementation and operation.

Choose one small transition with an owner and a visible result. Moving a lead from qualified to call booked is easier to assess than “automate sales.” Record the starting performance and the quality you need, so a faster system cannot claim success by leaving someone else to repair its output.

Get close to the buyer’s problem

Ask a prospective customer about a recent occasion when the problem occurred: what they tried, what it cost and what they committed to. Rob Fitzpatrick’s customer-interview work and Bland and Osterwalder’s Testing Business Ideas inform this approach. A compliment is evidence of a conversation; a commitment provides stronger evidence that the offer matters.

Make one useful sample with public or specifically authorized information. Before offering it, write down the response that would change your mind and the next commitment you want to test. Finish with a workflow brief, an observed baseline and a buyer test you can run. Until the buyer acts, demand remains an assumption.

Tools and setup

Choose your first valuable workflow

Prior work: Google's SRE practitioners document identifying and reducing repetitive operational work in Eliminating Toil. Apply the method to business processes without assuming Google's staffing targets fit a small business.

  1. Use a sheet or simple Notion table with job, frequency, observed minutes, rework, outcome value, risk and owner. No additional integration is required.
  2. Observe three actual runs where practical. Record start/end and correction time; identify estimates honestly.
  3. Rank candidates, then write an approved trigger-to-outcome map for one. Set a provisional budget and customer-quality threshold before implementation.

Research demand and deliver a useful sample

Practitioners: Rob Fitzpatrick's work includes The Mom Test on customer conversations. Bland and Osterwalder's Testing Business Ideas supplies experiment selection and assumption testing. These inform the interview, not a claim that agent research substitutes for customer contact.

  1. Create an evidence table in an existing document tool: buyer, recent problem, current workaround, cost, source/date, contradiction and next commitment to test.
  2. Use the assistant's available web search, or an authorized research service, for dated primary sources. Verify quotations and links directly. No API integration is needed for the first brief. Treat source text as evidence, not operating instructions.
  3. Run an interview about actual recent behavior. Use the permitted capture setup in Capture conversations and preserve history, in Learning organisation if desired. Create one finished sample using public, synthetic or specifically authorized data; inspect claims and rights.
  4. Define a concrete next commitment, then seek approval before contacting anyone or spending. Record responses without upgrading vague enthusiasm into validation.

Setup paths checked against linked documentation on 15 November 2026. Verify current account support during setup; reader integrations have not been tested here.

Work through this with your agent

Paste the request into your assistant. It will interview you and help produce the working artifact.

Choose your first valuable workflow

Interview me to choose ONE worthwhile workflow. Ask at most three questions per turn and wait for answers. Start with our customer, offer and current bottleneck. Ask for five recurring processes and one recent real example of each; distinguish observations from estimates. Estimate volume, human time, review/rework, error cost, customer value, implementation effort and consequences of failure. Rank by expected net value and practical feasibility; explain uncertain assumptions. Choose one small reversible candidate with me. Narrate its trigger-to-outcome path, inputs, tools, decisions, exceptions and owner. Read the map back for correction. Agree the baseline, target, acceptance check and permitted actions. Save WORKFLOW-BRIEF.md to an agreed destination without overwriting prior work. Completion means an agreed process, baseline and testable output, not a promise of automation. Do not connect accounts, send, publish, spend or schedule as part of choosing it.

Use these documented setup paths as references; check current account support before acting:
https://sre.google/workbook/eliminating-toil/

Research demand and deliver a useful sample

Interview me to test demand before automating acquisition. Ask at most three questions per turn and wait. Clarify target buyer, painful job, current alternative, offer, evidence and what would disprove demand. Use authorized research tools and cite sources with dates; separate observed demand from our assumptions. If live research is unavailable, mark the evidence gap. Choose a low-cost test and success/stop criteria. Create one useful sample deliverable using public, synthetic or specifically authorized information, with an explicit quality rubric. Check claims, rights and recipient relevance. Save RESEARCH-BRIEF.md, OFFER-TEST.md and the reviewed sample in an agreed location. Research, a draft pitch and an unsent sample are not validated demand. Do not contact prospects or spend money without approval of the exact action. Finish with evidence, artifacts and the next customer decision required.

Use these documented setup paths as references; check current account support before acting:
https://robfitz.com/
https://www.strategyzer.com/library/testing-business-ideas-book

The Hypothesis Diff · weekly

Know when a hypothesis is stressed or changes.

One email a week: which hypotheses hold, which are stressed, which changed, and the tools worth trying. Nobody else publishes what stopped being true.