The new rules of context engineering for Claude 5 generation models
TL;DR Highlight
Cutting over 80% of Claude Code's system prompt caused no measurable loss on coding evals.
Who Should Read
A developer whose bloated CLAUDE.md and Skills make Claude Code feel like its rules conflict with each other.
Core Mechanics
- Over 80% of system prompt removed
- Over 80% of Claude Code's system prompt was removed for Claude Opus 5 and Claude Fable 5.
- There was reportedly no measurable loss on coding evaluations.
- Rules were conflicting
- Internal transcripts showed "leave documentation as appropriate" and "DO NOT add comments" in the same request.
- Claude has to think harder when instructions overlap or conflict.
- Judgement instead of rules
- The old prompt said to default to no comments, one short line max.
- The new one says to write code that reads like the surrounding code, matching comment density, naming, and idiom.
- Interfaces instead of examples
- With the newest models, examples constrain exploration to that space.
- Design expressive tool parameters instead, such as a status enum of pending, in_progress, completed.
- Progressive disclosure over upfront loading
Community Reactions
- Agreement that long prompts are baroque
- One commenter said just talking to the agent beats long instructions.
- Another said that with good tools and a semi-clear request, the system prompt is nearly worthless.
- Experience turning auto-memory off
- One commenter disabled auto-memory and shaped CLAUDE.md, skills and docs by hand, getting more predictable results.
- They said agents write too much into memory and trim it badly.
- Others suspected vendor lock-in.
- Pushback about missing evidence
- Commenters said there is no proof this is better on real-world code and it may only work on Anthropic's codebases.
- One reported more mistakes and token use with Opus 5, and another said Fable 5 is harder to correct when wrong.
- Practical tips and experiences
- One commenter told Fable 5 to use its own judgement on tests and it worked well.
- Another uses counter-examples instead of examples.
How to Apply
- Delete obvious and conflicting lines in CLAUDE.md
- Open CLAUDE.md today and delete anything visible from the file structure.
- Replace clashing rules like "no comments" with "match the surrounding code style".
- Spend the remaining tokens on codebase gotchas, such as keeping types in one file.
- Split verification and review into a skill
- Move long verification steps from CLAUDE.md into a verification skill file.
- Leave only a one-line reference to it in CLAUDE.md.
- Replace tool examples with parameter design
- Remove usage examples from tool descriptions in your own agent.
- Design parameters, such as enums, so the usage is self-evident.
- State usage once in the tool description, not in the system prompt.
- Pass specs as code
- For long tasks, @-mention a test suite or HTML mockup instead of a prose description.
- Compare results to see the effect yourself.
Try It
# Paste into Claude Code as-is (the article's principles as a prompt, a simple version of claude doctor)
Below is my CLAUDE.md.
{{CLAUDE.md contents}}
Review it using these criteria.
1. Mark obvious content that can be learned from the file system or repo for deletion.
2. Find conflicting rules and merge them. Where possible, replace rules with judgement criteria like "write code like the surrounding code".
3. Keep only codebase gotchas.
4. Move long, situational guidance (verification, review, etc.) into separate skill files and leave a one-line reference in CLAUDE.md.
Show me the revised file and a list of what you removed.
Next step: also run the /doctor command in Claude Code and compare the results.Terminology
Related Papers
AI is a tool and using it well takes skill
A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome
Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting
Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.
GPT‑6 and Intelligent UI for everyone
GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.
Base Models Can Reason By Taking a Cue From Training Data
Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.
Labels Override Definitions in Jev-Style Typed Decision Models
Renaming options to A/B raised routing accuracy by +0.1511 over meaningful labels.
How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs
Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output