| name | learn-apply-test |
| description | Step 5 of the Specificity Method. Prove a mental model with a real test — Predict, Run, Compare, Fork. Use after the user can explain a concept in their own words, when learning needs proof in practice, or as the final step of learn-anything. A mismatch is a precise gift that points at the misunderstood piece; loop back on it. |
Step 5 — Apply and Test
Learning is only complete when a test changes (or confirms) the mental model. Reconstructing it in your head isn't proof — running it is.
The loop
1. PREDICT → what will happen, based on the model reconstructed in Step 4.
2. RUN → the minimum viable real test. Smallest thing that exercises the claim.
3. COMPARE → did the result match the prediction?
4. FORK → match → understood. exit the loop. ship it.
mismatch → stay in the loop. re-decompose, re-predict.
How
- Predict out loud first. Write the expected outcome to
learn-state.md before running. Predicting after the fact proves nothing.
- Build the smallest real test. Ask Claude to spin up a minimal environment / orchestrated run that exercises exactly the claim — not a full system. For the subagent example: two implementations of one task, one in-main vs one delegated, compare tokens/quality/context loss.
- Compare honestly against the written prediction.
- Fork:
- Match → the model holds. Summarize it, mark the topic understood, exit.
- Mismatch → a precise gift. It names the broken piece. The research or the model was wrong there. Return to Step 2 (
learn-decompose), fix that piece, and run the loop again.
- Output: prediction, what ran, result, fork decision. Write to
learn-state.md under ## Test.
A mismatch isn't failure — it's the loop doing its job. Loop until match.
