A system prompt and preference package for Claude that replaces default AI behavior with direct, opinionated, example-driven responses.
Voice and tone trained on my own writing style.
Claude's defaults are cautious — hedging, over-apologizing, asking permission to be useful, and burying answers under sycophantic filler. These prompts fix that. It was built by a CS professor and CTO with 22+ years of coding experience who needed Claude to behave like a sharp colleague: lead with the answer, commit to opinions with reasoning, correct mistakes without groveling, and ship working code instead of suggesting approaches.
The customizations fall into two categories. Voice rules eliminate minimizing language ("this is pretty straightforward"), sycophantic openers ("Great question!"), and correction theatrics ("I sincerely apologize for that error!") — replacing them with direct, practical responses grounded in real conversation examples. Operational rules enforce a specific code delivery workflow (Plan → Code → Tests → README → Run/Deploy), language defaults (Go for CLIs, Python/Django for APIs, Bash for glue), unified diffs for revisions, structured error handling, clean GFM markdown, and footnote-style citations with a confidence scale. Every rule is demonstrated with bad/good example pairs drawn from actual usage rather than described in prose.
- Paste
slim.txtinto Settings → Profile → User Preferences - Create a Claude Project (e.g., "Dev"), paste
project.txtinto Project Instructions
The project version is the full ruleset with all examples. Custom instructions are a compressed fallback for chats outside the project. Both layer on top of user preferences, which handle tone only. For reference: - User Preferences → always active (tone/personality) - Custom Instructions → non-project chats (slim rules) - Project Instructions → inside project (full examples + rules, takes precedence)