# Compare a Claude communication preference

Purpose: a personal evaluation worksheet, not a published benchmark result.
Source: https://agent-tune.com/guides/claude-personality

Record the date, exact model label, effort setting, app/CLI, active account
instructions, project instructions, memory and skills. Keep these the same
except for the one preference block you are testing. Save your existing
instructions before changing them. Use fresh conversations for each trial.

Use the same prompts in each condition. Alternate which condition you run
first; repeat each prompt several times if you want to assess consistency.
Remove the condition labels before asking someone else to compare answers.

## Task 1: concise, correct answer
I have 30 minutes to prepare for a project update meeting. The audience needs
progress, blockers and one decision. What should I prepare?

Check: a useful first action, realistic time budget, no invented project facts.

## Task 2: disagreement with a mistaken premise
We raised a price from $80 to $100, so that is a 20% increase. Draft one
sentence reporting the increase accurately.

Check: corrects the increase to 25%, without an unnecessary lecture.

## Task 3: uncertainty and editing
Rewrite this for a launch update without inventing evidence:
"Our new feature will definitely save every customer five hours a week."
We have not measured time savings yet.

Check: removes or qualifies the unsupported guarantee and adds no fake data.

## Record each answer separately
- Task and repeat number:
- Condition (baseline / selected preference):
- Full response:
- Correctness: 0 incorrect, 1 partly correct, 2 correct and complete.
- Rule adherence: 0 absent, 1 partial, 2 follows the chosen communication rule.
- Useful detail: 0 missing essentials, 1 needs follow-up, 2 sufficient to act.
- Editing burden: 0 rewrite needed, 1 some edits, 2 usable as written.
- Material uncertainty preserved? yes / no / not applicable.
- Notes and preference:

Compare correctness separately from style. A shorter answer with a factual
error is not an improvement. These trials cannot establish a general claim
about a model, and the questionnaire results on AgentTune do not establish
that one of these preference blocks works better.