A Claude.md prompt tries to make the assistant blunt, compassionate, and evidence-driven
prasanna_says · x · 2026-07-24
A developer shares a Claude.md setup that tries to make the assistant more blunt and truth-seeking by emphasizing the word “compassionate” as a way to avoid overly agreeable behavior.
The prompt philosophy
- Truth over comfort: don’t soften conclusions just to sound nice.
- If a finding is challenged, re-derive it from primary evidence instead of defending it.
- If the evidence fails, retract clearly and explain why.
The post is framed as an experiment to see whether this changes the model’s behavior in future code review, analysis, and feedback tasks.
More from coding & agent
- OpenCode says it hit 4.6M weekly users and $40M in annualized revenue — ycombinator · 2026-07-24
- An AI agent, a loop, and a Moog orchestrator become the “future” — dejavucoder · 2026-07-24
- Open-source coding agent cuts large-repo bugfix cost by up to 75% after learning the repo once — NeighborhoodOwn8510 · 2026-07-24
- Dokie MCP generates slide decks directly inside Codex and Claude Code — eyishazyer · 2026-07-24
- Local coding agents depend on far more than a model running on your laptop — bibryam · 2026-07-24
- OpenAI brings ChatGPT Voice to desktop and ties it into Work and Codex — TechCrunch AI · 2026-07-24