disler/fixing-smartass-opus-5

What it solves

Opus 5 is one of the smartest models available, but its default communication style wastes tokens and time. It buries answers under multiple headers, uses repetitive phrases like “load-bearing” and “worth stating plainly,” adds flattery, and even stamps co-author credit on commits. Engineers often try to fix this by tweaking individual user prompts, which only helps one task at a time. The real leverage is in the system prompt, which applies to every turn of every session. This project provides a carefully written system prompt that turns a verbose, scope-creeping assistant into a concise, precise engineering partner.

How it works

The entire product is a single markdown file, sr_opus_5_system_prompt.md, that you append to your coding agent's system prompt using the --append-system-prompt-file flag. The prompt is organized into four instruction sections plus concrete examples:

• Positive and negative patterns: A do-this/never-do-this list. It bans phrases like “load-bearing” and em dash chains, and reinforces concise, direct communication habits on every response.
• Reference points: The model labels its output with short codes — F1 for findings, D1 for decisions, O1 for options, R1 for risks, Q1 for questions, A1 for actions. Your follow-ups become near-zero-token commands like keep D1, reject O2, answer Q1.
• Hard operational boundaries: Scope control that blocks unrequested refactors, adjacent cleanup, completion claims without evidence, and co-author credits in commits.
• Aliases: Two-to-three letter commands that expand into full instructions, e.g. scr (simplify, compress, repeat), eli (explain like I'm 18), foc (focus on the one thing), and ref (rewrite with reference points).

There are also examples built from real model outputs — a “To do” and a “Not to do” — because models pattern-match examples more strongly than rules.

The repo includes a comparison loop. Using just commands and herdr, you can launch two Claude Code panes side by side: one with the stock model, one with the system prompt appended. Both get the same user prompt (summarizing a long blog post), and you watch response length, output tokens, and wall-clock time diverge in real time. The same system prompt also works with the Pi coding agent, proving it's portable across harnesses.

Who it's for

Mid to senior engineers who run frontier coding agents (like Claude Code, Pi, or any harness with an appendable system prompt) and want to cut costs and speed up iteration without sacrificing output quality. It's for anyone tired of tuning user prompts one task at a time and ready to fix the communication channel at the system level.

Highlights

• Single-file, no build or dependencies — just plain markdown and a --append-system-prompt-file flag.
• Explicit banned-phrase list and positive behavior patterns to reinforce concise, no-bs responses.
• Reference codes turn follow-ups into near-zero-token commands.
• Hard boundaries stop scope creep, unrequested cleanup, and co-author credits.
• Aliases like scr, eli, foc, ref act as bash aliases for conversation.
• In-context examples distill tone better than rules alone.
• A built-in compare loop (just compare <name>) shows the delta side by side with the stock model.
• The same system prompt works across Claude Code and Pi, proving portability.