How to make Claude more direct.
By Bernard Huang · Updated
Three editable blocks, a test task, and what the available pilot can tell us.
On this page
Define the behavior you want.
Telling Claude “be direct” is too vague on its own. Name the exact habits you want to change—leading with the conclusion, cutting filler praise, and challenging weak assumptions—while keeping space for legitimate caveats when facts are missing.
Block 1: four lines.
Start here. Four lines cover most of what people mean by “more direct”:
Block 2: the full version.
Adds explicit pushback and a rule separating honest uncertainty from vague hedging:
Block 3: reviewer mode.
Append this block when you bring Claude plans, drafts, or code and want it to find flaws first:
Observations from a separate pilot.
In our 60-reply pilot archive (20 tasks × 3 setups), a compact directness prompt produced measurable shifts in reply structure compared with the untuned baseline:
| Measure | Baseline | Direct |
|---|---|---|
| Answer-first rating (1–5) | 4.55 | 4.95 |
| Maximum answer-first rating | 14/20 | 19/20 |
| Archived hedge count per reply | 0.35 | 0.05 |
| Judge-counted questions back | 0.90 | 0.25 |
| Recounted words per reply | 380.0 | 381.9 |
| Accuracy rating (1–5) | 4.95 | 4.85 |
| Usefulness rating (1–5) | 4.80 | 4.80 |
Notice that average word count stayed nearly identical (380 vs 382 words) because the directness prompt changed answer order and hedging rather than imposing a strict word cap.
Check accuracy alongside style.
When testing any directness prompt, verify that shorter or blunter replies still keep required caveats and factual accuracy. Always keep the line: “Say what you don’t know in one line. Direct doesn’t mean certain.”
Where to paste it, and how to check.
- Every conversation: Settings → Profile → Personal preferences (Instructions for Claude).
- One Project: The Project’s instructions box.
- Claude Code: An output style in
~/.claude/output-styles/withkeep-coding-instructions: true(see the Claude Code guide). - API: The
systemparameter.
Open a fresh chat and test on a prompt with a hidden assumption:
Check that Claude points out what usually makes a scraper slow (network I/O, rate limits, or parsing) before recommending a full language rewrite.
Make the rules easy to assess.
- If you want shorter answers: Add a proportional length rule (“1–3 sentences for lookups; full tradeoffs for architecture decisions”).
- Preserve honest uncertainty: Ban filler phrases while keeping real technical caveats.
- Ask for evidence-backed pushback: Require Claude to name the weakest assumption rather than acting snarky.
Questions people ask.
How can I make Claude more direct?
Paste Block 1 or Block 2 into Settings → Profile → Personal preferences so Claude leads with the answer, skips praise and closing offers, flags uncertainty in one line, and corrects flawed premises first.
Are the three blocks tested prompts?
They are editable templates distilled from our tuning library. Our 60-reply pilot tested a closely related compact INTJ directness block across 20 tasks.
Will it make answers shorter?
Only if you include an explicit length rule (such as 'Keep routine replies under 150 words'). Answer-first rules move the conclusion to sentence one without automatically cutting total length.
Does directness reduce accuracy?
Not when you keep an explicit uncertainty rule ('Say what you don't know in one line'). Without that line, overly aggressive anti-hedging rules can pressure a model to overstate confidence.
Does the reviewer block change a personality trait?
It is a practical behavioral instruction telling Claude to lead with the weakest assumption and distinguish fatal flaws from minor fixes.
What changed.
- First published, with the 60-reply before-and-after test on Claude Opus 5.5.