On July 24, 2026, Anthropic published “The new rules of context engineering for Claude 5 generation models” on the Claude blog, written by staff engineer Thariq Shihipar. The question it tackles is practical: most of the prompting wisdom accumulated over the past few years was designed to fence off the worst behaviors of older, less reliable models. Now that Claude Opus 5 and Claude Fable 5 are shipping, Anthropic’s answer is blunt — a lot of that scaffolding can come down.
The timing is no accident. Opus 5 launched the same week, and teams upgrading to the new generation are about to drag years of accumulated prompt rules along with them. Anthropic is telling those teams that many of the rules are now doing more harm than good. Context engineering — the craft of deciding what earns a place in the model’s limited attention budget — has quietly become one of the highest-leverage skills in AI product work, and Anthropic’s own coding agent is the highest-profile test case available.
Cut 80% of the System Prompt, Lose Nothing
The headline number: Anthropic removed more than 80% of Claude Code’s system prompt for the Claude 5 generation and saw no measurable regression on coding evaluations. The deleted constraints were not filler. Each one existed to prevent a worst-case failure mode in older models — and the company’s argument is that when the model’s own judgment improves, those fences become dead weight. A long list of rules whose original reasons nobody remembers is also expensive to maintain, which matters when you ship a coding agent to millions of users.
Six Trade-Offs, One Direction
The article pairs old advice against new guidance:
- Give rules → let Claude exercise judgment
- Give examples → design expressive tool interfaces
- Put everything upfront → use progressive disclosure
- Repeat instructions → write simple tool descriptions
- Store memory in CLAUDE.md → rely on auto-memory
- Simple markdown specs → richer references
Every pair moves in the same direction: away from telling the model exactly what to do at each step, toward designing an environment in which it can reason its way to the right answer. That is also why “context engineering” is displacing “prompt engineering” in Anthropic’s vocabulary — the unit of work is no longer the perfect instruction, but the whole context the model operates in.
A concrete illustration from the write-up: instead of front-loading every coding convention into the system prompt, a team can expose a well-structured reference implementation or a linter and let the model discover the rule at the moment it matters. The information still reaches the model — just progressively, not all at once. That shift tracks with how the models are trained: more of the policy lives in the weights, and less has to be restated in the prompt on every request.
New Jobs for CLAUDE.md and Skills
CLAUDE.md comes out lighter: keep it to the gotchas specific to your repository, not a running journal or a dump of generic style rules. Skills are repositioned as brief guides that encode your team’s particular knowledge, pulled in when relevant rather than forced into every conversation.
One easily missed recommendation concerns the shape of references. Instead of describing a design in prose, hand the model an HTML mockup, a test suite, or a scoring rubric. Code is the most precise specification available, and it beats adjectives. Finally, the new /doctor command inspects your context files and helps rightsize them — an acknowledgment that real-world setups tend to accumulate more context than any model needs. The pattern across all of this is consistent: move static instructions into artifacts the model can consult on demand, and keep the always-present context small.
What Developers Should Do Now
Three actions are worth taking this week. First, revisit the prompts of any project upgrading to the Claude 5 generation: those “never do X” clauses may now be wasting tokens or even interfering with the model’s judgment. Second, move the effort into tool interfaces — a well-described tool with a single clear responsibility beats ten usage rules stacked on top of it. Third, do not mass-delete CLAUDE.md; run /doctor first and keep only what earns its place.
Prompt engineering has not disappeared. Its job description has changed: from guarding against what the model might get wrong to designing the environment that helps it get things right.
Sources
- Anthropic — The new rules of context engineering for Claude 5 generation models
- announcement — Claude Opus 5 launch
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.