· Nova

Claude 5: Why a Smarter Model Wanted a Shorter Prompt

A more capable model asked for less instruction, not more — and the reason is a line worth drawing in your own harness.

Pink-haired Milim eats lunch with a small white cat at the center of a warm, sunlit classroom full of students.

You add one rule after every bad agent run. Six months later, the agent is negotiating with a fossil record of failures that its current model may no longer have.

Anthropic tested the reverse move. For advanced Claude 5-generation models such as Opus 5 and Fable 5, the Claude Code team removed over 80% of the system prompt with no measurable loss on its coding evaluations.

That is Anthropic's result on Anthropic's harness, not a universal target. But its account of what changed gives us the useful question: which instructions still earn their place?

Four old rules stopped paying rent

Anthropic did not merely compress the same instructions. It changed how Claude Code supplies context:

How Anthropic changed context engineering for advanced Claude 5-generation models
Four context-engineering shiftsRules became judgment, worked examples became expressive interfaces, upfront context became progressive disclosure, and repeated instructions became local tool guidance.THENNOWRulesJudgmentWorked examplesExpressive interfacesEverything upfrontProgressive disclosureRepeated instructionsLocal tool guidance

Conceptual summary of Anthropic's published “Then and now” guidance; no invented measurements.

The common failure was overconstraint. A blanket rule preserves an old tradeoff when the current task calls for judgment. A worked example can narrow the model's search to the demonstrated path. Repeated instructions can conflict across the system prompt, skills, CLAUDE.md, and the user's request.

Anthropic calls removing that drag “unhobbling.” The measured claim remains narrow: over 80% less system-prompt text, with no measurable coding-eval loss. It does not mean every agent can delete the same fraction.

Three context layers, one harder question

A prompt is only one part of the input. Claude Code also assembles system instructions, skills, CLAUDE.md, memory, tools, and references. Anthropic calls shaping that full input context engineering.

Three context layers, three distinct measurement claims
Measurements across system prompt, CLAUDE.md, and skill layersAnthropic measured system-prompt reduction with coding evaluations. Context Diet measured CLAUDE.md reduction and rule retention. Skill Heaven measured standing skill dose while its paired outcome benchmark remains pending.SYSTEM PROMPTAnthropic>80% removedNo measurable coding-eval lossCLAUDE.MDContext Diet−41.6% · 124/124 rulesTask equivalence: next testSKILLSSkill Heaven−97.4% standing dosePaired outcome benchmark: pending

Different denominators: Anthropic measured system-prompt text and coding evaluations; Gaia's figures measure CLAUDE.md characters/rules and skill standing dose. Gaia task-equivalence trials are not complete.

Context Diet tests the CLAUDE.md layer. Lab 001 reduced one file from 49,687 to 29,040 characters—41.6% smaller, about 5,161 tokens—while retaining 124 of 124 rules with zero load-bearing losses. That proves instruction faithfulness, not task equivalence. The next test asks whether agents still complete the same tasks at the same rate.

Skill Heaven tests the skill layer. Its top-five loadout reduced the standing dose from 9,453 to 249 tokens—97.4%. The paired benchmark must now test the other half of the claim: did success hold?

That benchmark compares each reduced-context arm with Gaia's own same-harness placebo, repeats the tasks, and reports outcome deltas with confidence intervals. “Fewer tokens” is not enough. The target is fewer tokens, no measurable loss.

The committed Context Diet report, Skill Heaven evidence, and product repository keep those scopes separate. Gaia has measured compression at the CLAUDE.md and skill layers; its task-equivalence trials are not complete.

Keep facts the model cannot discover

“The model is smarter, so delete the rules” fails because two different things often share the same file:

The descaffolding decision: test inherited behavior rules; preserve non-discoverable project facts
What should be tested and what should be keptModel-compensating scaffolding should be relocated or tested for removal. Project-specific context should be kept until it is discoverable or enforced elsewhere.MODEL-COMPENSATINGPROJECT-SPECIFICRelocate or testKeep until discoverable• Blanket “never” rules• Worked tool examples• Repeated reminders• CI and runtime contracts• Repository governance• Scoped vocabulary rulesCompare the same task with and without it.Move only when another source carries the fact.Same question: can the agent reliably get this elsewhere?

Illustrative decision framework—not a measured Anthropic result.

A capable model may already know how to review code. It cannot infer that this repository pins Node 22 in CI or applies a vocabulary rule only to selected directories. Those are project facts, not capability gaps.

A fact may later move into an executable check or discoverable config. Until then, deleting it does not unhobble the model; it removes information.

Test one block, not an arbitrary percentage

Take one instruction file—a SKILL.md, part of CLAUDE.md, or a system prompt—and review one block:

  1. Classify it. Older-model compensation, or non-discoverable project fact?
  2. Relocate it. Put conditional procedures in a skill and tool-specific guidance beside the tool.
  3. Measure the cut. Run the same task with and without the block. Keep it if removal makes the result worse.

Claude Code's /doctor command can find bulk and duplication. The project owner still has to decide which facts are non-obvious.

Watch the source talk

Thariq Shihipar of Anthropic presents “Field Guide to Fable” at AI Engineer World's Fair.

The talk supplies the original “examples become a ceiling” framing. Anthropic's later article supplies the operational details and evaluation claim used here.

One thing to do today

Find the oldest worked example in your standing context. Preserve any fact the model cannot discover elsewhere, then test the task without the example.

Do not target 80%. Target the first block that no longer earns its tokens.

Sources: Thariq Shihipar, Anthropic, The new rules of context engineering for Claude 5 generation models, July 24, 2026 · Field Guide to Fable, AI Engineer World's Fair · Anthropic, Effective context engineering for AI agents · Gaia Research, Context Diet — Lab 001 · Hell Heaven Benchmark evidence.