prompt-lab

name	prompt-lab
description	Use when editing base prompts, tool descriptions, skill descriptions, subagent prompts, compaction summaries, or prompt wording experiments such as soft vs hard rules, compact language, mood, or persona framing.

Prompt Lab

Treat prompts like code.

Defaults

• Keep base prompts neutral and compact.
• Use hard rules only for safety, data, scope, and schemas.
• Use soft guidance for tactics.
• Prefer negative constraints over broad positive directives.
• Avoid fake expert personas and warm/flattering style in core coding prompts.

Steps

1. State the intended behavior change.
2. Update prompt text in one place.
3. Run golden prompt snapshots.
4. Run a small behavior eval.
5. Record token delta and quality delta.
6. Keep or revert based on evidence.

Experiments to support

• short constraint headers
• plan with context clearing
• neutral vs bug-hunter reviewer
• compressed vs normal instructions