Skip to main content
Unlisted page
This page is unlisted. Search engines will not index it, and only users having a direct link can access it.

Worksheet — Module 2 lab

Optional. Copy this file, rename it (e.g., worksheet-2026-04-24.md), and fill it in as you work through the lab. If you'd rather use your own notes format, ignore this.


Step 1: Cold read​

Date/time: _____

First impressions after reading the conversation without diagnosing:

  • What's the developer trying to build?


  • How does the conversation feel as it progresses? Where (roughly) does it start to feel off?


  • Any initial gut calls on what's going wrong? (Don't verify yet — just record the instinct.)



Step 2: Diagnostic pass​

For each turn where the model's response feels off, note the turn number, what you think is wrong, and an initial failure-type guess. Don't limit yourself to a specific count. Some off-sounding turns will turn out to be fine.

TurnWhat looks offInitial diagnosisEvidence (scroll-back result)

Diagnosis key: forgot | never had it | correction loop | over-contextualization | actually fine | unsure


Step 3: Pick three to fix​

Pick a mix of failure types rather than three of the same kind.

Fix 1​

  • Turn: _____
  • What's wrong (one line): _____
  • Failure type: _____
  • Single-move fix: _____
  • Prediction (one sentence — what should the output look like after the fix?):

Fix 2​

  • Turn: _____
  • What's wrong (one line): _____
  • Failure type: _____
  • Single-move fix: _____
  • Prediction:

Fix 3​

  • Turn: _____
  • What's wrong (one line): _____
  • Failure type: _____
  • Single-move fix: _____
  • Prediction:

Step 4: Run the fixes​

For each fix, run it in a new Claude conversation and compare against your prediction.

Fix 1 — result​

  • Prediction held? yes / partially / no
  • What the output actually did:
  • If the prediction missed: what did you miss about the diagnosis?

Fix 2 — result​

  • Prediction held? yes / partially / no
  • What the output actually did:
  • If the prediction missed: what did you miss about the diagnosis?

Fix 3 — result​

  • Prediction held? yes / partially / no
  • What the output actually did:
  • If the prediction missed: what did you miss about the diagnosis?

Step 5: Calibration check​

Looking back at your Step 2 notes: which turns did you initially flag as off that you now think were actually fine given the context?

  • Turn _____: actually fine because _____
  • Turn _____: actually fine because _____

Any turns you didn't flag initially but now think were off?

  • Turn _____: missed because _____

Takeaways​

Two or three sentences on what you noticed. Examples of useful observations:

  • The failure type you found hardest to diagnose and why
  • A fix that worked better than you expected, or worse
  • A pattern from this lab you expect to see in your own real conversations


If you did the extension​

Three fixes for a single "Claude forgot" case:

ApproachHow it performedWhen you'd reach for it
In-turn restatement
Summary-then-ask
Fresh conversation with distilled context

Which one worked best for this specific failure, and what does that suggest about when each approach is the right move?