Skip to lesson
中文
The collection/AN OPEN STUDY ROUTE

WEEK 13 / 24 · Learn, then operate

Knowledge into action

Turn knowledge into action: make a prediction and run a falsifiable experiment

10 core hours + 2 optionalREAD · BUILD · PRACTISE
Begin the lesson ↓
Knowledge into actionTHE PERSONAL FDE FIELD LIBRARY

The question: Which decision, action or workflow changed because of something you learned this week?

Before you begin: Have traceable sources and evidence cards, and distinguish facts, inferences and hypotheses.

FIG. 13.1THIS WEEK’S REASONING PATH
  1. 01Evidence
  2. 02Prediction
  3. 03One-variable trial
  4. 04Adopt or reject
Teaching diagram · Derived from this week’s lab; a suggested process, not observed results.

Concepts and learning objectives

Give knowledge an exit into practice. Turn a claim into an observable prediction: under which conditions will changing which variable affect which result? Include counterexamples, costs and stopping conditions. Improvement after a change does not prove causation; a small comparison helps resist wishful thinking.

Choose a low-risk technical question: whether structured input reduces rework, short handoffs shorten recovery or source filters reduce obsolete citations. Health, investment, employment and other consequential decisions are outside these experiments. Change one principal variable and keep failed results.

Reading and selected sections

Priority Source and selected sections Budget
Optional R47 · OpenAI — Harness engineering — Extract one small, transferable hypothesis about feedback or constraints. 35 min
Required review R42 · Anthropic — Demystifying evals for AI agents — Convert an expected improvement into outcome checks. 35 min
Required selections R45 · Anthropic — How AI assistance impacts the formation of coding skills — Task delivery versus learning; respect the study's limited scope. 35 min

Studio lab

  1. Pick an evidence-backed claim relevant to an actual friction. Write your own explanation and prediction before asking the model. Include conditions under which it may fail.
  2. Design a 30–60-minute experiment with a fixed task, baseline and decision rule. Count preparation, execution and review, not only generation speed.
  3. Run baseline and changed versions and retain all results. Where model variation matters, repeat modestly and record the spread.
  4. Decide to adopt, reject or gather evidence. Adoption should change code, a checklist or a default, rather than produce an “I should remember” note.
  5. Attach the change to a natural trigger for a real task, then plan a later test on a fresh example. Avoid an ever-growing reminder system.

Deliverables and acceptance

Deliverable Evidence of completion
Action card Source, explanation, prediction, counterexample, experiment, result and actual change.
Reproducible experiment Baseline, intervention and checking conditions; failures retained.
Behavioural decision A change is implemented or explicitly rejected for a reason.

Independent exercise: Propose a second explanation for the result and design a small check to distinguish it. Another model's agreement is not a substitute.

Privacy and safety: Change behaviour in isolated copies using synthetic data. Sending messages, deleting, paying or changing important calendars are not low-risk experiments.

Everyday use: Apply the rule to a task that recurs this week. Observe hesitation, search and rework, including the possibility that the old habit was already sufficient.

Pass or revisit: A claim leads to an observable decision, including rejection. If only summaries accumulate, stop expanding the source library.

A reusable collaboration prompt

Turn one sourced technical claim into a testable hypothesis. Design a low-risk experiment with a fixed baseline and stopping condition. Change one main variable, retain failures and alternatives, and allow adoption, rejection or insufficient evidence.

Sources and verification scope

Reference IDs and scope follow the supplied manuscript, baseline 2026-09-27. External authenticated labs were not all completed; recheck versions and access before implementation.

  1. R42
    Anthropic — Demystifying evals for AI agents ↗

    9 January 2026; outcome evaluation and process inspection.

  2. R45
    Anthropic — How AI assistance impacts the formation of coding skills ↗

    Controlled study with bounded sample and tasks; not a claim about all AI-assisted learning.

  3. R47
    OpenAI — Harness engineering ↗

    11 February 2026; repositories, feedback and constraints, without assuming internal returns transfer.