THE NEW RULES Anthropic deleted 80% of Claude's system prompt

@vibewithkevin

8.9K followers · usually 650 views a post

Views

15.2K

Likes

150

Comments

Selling · Course / Info

The new rules of context engineering

Why it worked

The post provides valuable, actionable insights into optimizing prompts for AI models like Claude, which is a relevant and trending topic for developers and AI enthusiasts. The "then vs. now" format clearly illustrates the evolution of best practices, making the information easy to digest and apply.

Summary

This post explains how Anthropic reduced Claude's system prompt by 80% without measurable loss in coding evaluations. It details five shifts in prompt engineering, moving from rigid rules to more nuanced judgement, examples, and context-aware prompts.

Structure

1Introduction: Claude's system prompt reduced
2Why it worked: Conflicting rules caused inefficiency
3Shift 1: Rules to judgement
4Shift 2: Examples to interface design
5Shift 3: Front-loading to on-demand skills
6Shift 4: Hotkeys to auto-memory
7Shift 5: Markdown specs to richer references
8

Original caption

Thariq's post is the closest thing we have to official guidance on context engineering for @Claude opus 5. the reason they could cut that much is the part worth reading. they read their own transcripts and found single requests carrying "leave documentation as appropriate" and "DO NOT add comments" at the same time, from the system prompt, a skill, and the user, all clashing. the model sorts it out, but it burns thinking to do it. five shifts in the post. 1. rules give way to judgement. 2. tool examples lose to good interface design. 3. everything-upfront becomes progressive disclosure. 4. auto-memory takes over the manual # saves. 5. spec can now be a test suite, an html artifact, or a rubric. Going to go cut mine tonight. #techtok #ai #coding #claudeai #aiagents

More from @vibewithkevin

Their other posts in the index, biggest breakout first.

On-screen text

OPUS 5 · JULY 2026 01 / 09 THE NEW RULES Anthropic deleted 80% of Claude's system prompt No measurable loss on their coding evals. Thariq wrote up what changed for Opus 5, and most of it lands on the CLAUDE.md in your repo right now. SYSTEM PROMPT Default to writing no comments. Never write multi paragraph docstrings. Don't create planning or analysis docs. Detailed code review steps, inline. Tool instructions, repeated up top. Write code that reads like the surrounding code. @trq212 claude.com/blog Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin SWIPE → OPUS 5 · JULY 2026 02 / 09 WHY IT WORKED Your rules were fighting each other They read transcripts of their own internal use and found single requests carrying both of these at once, from the system prompt, a skill, and the user. Claude can sort that out. It just has to spend the thinking to do it first. ONE REQUEST leave documentation as appropriate VS DO NOT add comments Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin SWIPE → OPUS 5 · JULY 2026 03 / 09 SHIFT 01 Rules out, judgement in The hard bans existed because older models got comments wrong often enough to be worth the tradeoff. That stopped being true, so the rule became a description of taste. THEN → NOW THEN In code: default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks. NOW Write code that reads like the surrounding code: match its comment density, naming, and idiom. Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin SWIPE → OPUS 5 · JULY 2026 04 / 09 SHIFT 02 Examples box the model in Rule one for tool use used to be show it an example. On the newest models that narrows where it will look. Put the effort into parameters instead. THEN → NOW THEN Give Claude examples of how to call every tool. NOW Design the interface. A status enum of pending, in_progress, completed teaches the tool by itself. Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin SWIPE → OPUS 5 · JULY 2026 05 / 09 SHIFT 03 Stop front-loading everything Code review and verification left the system prompt and became skills Claude calls when it needs them. Some tools now hide behind a search until they get used. Your CLAUDE.md can be a tree instead of a wall. THEN → NOW THEN One CLAUDE.md holding every practice you might ever hit. NOW A short CLAUDE.md that points at skills, loaded at the moment they apply. Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin SWIPE → OPUS 5 · JULY 2026 06 / 09 SHIFT 04 The # hotkey is not the move anymore Hand-saving memories into CLAUDE.md was the old story. Claude now writes its own memories for the work and for you, which means one less thing your context file has to carry. THEN → NOW THEN Hit # to append what you learned to CLAUDE.md. NOW Auto-memory saves what is relevant to the work and to you. Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin SWIPE → OPUS 5 · JULY 2026 07 / 09 SHIFT 05 A spec does not have to be markdown Claude handles richer references now. A test suite is a spec. So is an HTML artifact, a function in another codebase worth porting, or a rubric you hand a verifier agent. THEN → NOW THEN plan.md and spec.md, written in prose. NOW Test suites, HTML artifacts, real code to port, rubrics for verifier agents. Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin KEEP BUILDING OPUS 5 · JULY 2026 08 / 09 DO THIS TONIGHT Rightsize the files you already have Keep CLAUDE.md light on what the repo is and heavy on the gotchas Claude cannot see from the file tree. Keep skills opinionated, and split the long ones across files. TONIGHT ☐ Cut rules written for a worst case that stopped happening ☐ Move the situational guidance out into skills ☐ Delete anything Claude can read off your file tree ☐ Run /doctor, which ships to do this pass for you Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin SWIPE → OPUS 5 · JULY 2026 09 / 09 SCREENSHOT THIS Your CLAUDE.md is probably too long The team that ships the model cut 80% of theirs and the evals did not move. Worth an hour on yours. SAVE. SHARE. FOLLOW. Thariq (@trq212) · The new rules of context engineering · Jul 2026 00000000 @vibewithkevin KEEP BUILDING