← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Sourceupdated Sat Jul 25 2026 08:00:00 GMT+0800 (Philippine Standard Time)

The New Rules of Context Engineering for Claude 5 (Thariq Shihipar)

claude-codeanthropiccontext-engineeringskillsclaude-mdcapability-overhangprompt-engineeringprogressive-disclosuremythos-classfable-5opus-5

The New Rules of Context Engineering for Claude 5 (Thariq Shihipar)

Thariq Shihipar (Member of Technical Staff, Anthropic, working on Claude Code) on the official Anthropic blog, published July 24, 2026. The blog counterpart to his AI Engineer World's Fair talk — the WF talk gave the personal narrative and the "why" (Capability Overhang); this blog gives the operational "how" as a set of named pattern shifts for authoring system prompts, skills, and CLAUDE.md files for the Claude 5 generation (Opus 5 / Fable 5).

Primary finding

We removed over 80% of Claude Code's system prompt for Claude Opus 5 and Claude Fable 5 models "with no measurable loss on our coding evaluations."

The core mechanism the blog articulates: newer models require less explicit constraint and can leverage better judgment to interpret context. The old rules existed because older models made poor decisions without guardrails; on the new class, the same rules block the model from reaching its own imagination. See Capability Overhang for the vault's concept page on this.

Shihipar's headline diagnosis: "We found that we were overconstraining Claude Code, both through our system prompt and in our CLAUDE.md files and skills."

Concrete before/after — the reduction pattern

The blog gives a canonical example of what "rightsize by deletion" looks like inside a system prompt:

Old (constraint-heavy):

"In code: default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks — one short line max. Don't create planning, decision, or analysis documents unless the user asks for them."

New (judgment-first):

"Write code that reads like the surrounding code: match its comment density, naming, and idiom."

The new line does the same job with one thought and no negations, and it generalises across languages and repos the model will encounter.

The six pattern shifts

The load-bearing structural claim of the blog. Each shift replaces a constraint-era practice with a judgment-era one.

# Old (constraint era) New (Claude 5 era) The shift
1 Give Claude explicit rules to avoid worst-case scenarios Let Claude use contextual judgment Rules → Judgment
2 Provide tool usage examples to constrain behaviour Design expressive tool interfaces (parameters, enumerations, hints) Examples → Interface design
3 Put all detailed guidance in the system prompt Load context selectively via skills, deferred tool definitions, and ToolSearch Upfront context → Progressive disclosure
4 Repeat instructions in system prompt and tool descriptions Put tool usage instructions only in tool descriptions Repetition → Simple descriptions
5 Users manually save memories to CLAUDE.md files Claude automatically saves relevant memories to work context Manual memory → Auto-memory
6 Store specs as simple markdown files Use HTML artifacts, code samples, detailed test suites, or function implementations as references Simple specs → Rich references

Direct-quote anchor for shift #2 (examples now constrain rather than help): examples "constrain them to a certain exploration space." Direct-quote anchor for shift #3: avoid "making these a central repository for every known practice." Direct-quote anchor for shift #6: "A spec may also be a detailed test suite, or a function in a different codebase that Claude might port."

Where each rule lands — by artefact

The blog is unusually specific about which rule applies to which surface a practitioner authors:

System prompt

  • Heavily tied to product context — tell Claude what product it operates in.
  • Spend most of the authoring effort here when building a custom agent.
  • Avoid stating "the obvious" things Claude should know by looking at the file system.

CLAUDE.md files

  • Keep lightweight; briefly describe the repository's purpose.
  • Focus tokens on non-obvious gotchas and exceptions.
  • Use progressive disclosure for complex instructions — split into skills.
  • Create separate skills for verification, validation, or specialised guidance rather than folding them into CLAUDE.md.

Skills

  • Lightweight guides for information retrieval.
  • Avoid overconstrained instructions except in genuinely critical areas.
  • Divide long skills across multiple files.
  • Best when encoding team-specific opinions, knowledge, or best practices.

References (@mentions)

  • Include in-depth specs, mockups, or entire codebases.
  • Prefer code references over descriptions.
  • "An HTML mockup of a design will generally produce better results than a description of the design or a screenshot."

Patterns to avoid (blog's own list)

  • Conflicting instructions across system prompt, skills, and user request → unnecessary cognitive load on the model.
  • Over-constraining through redundant rules.
  • Central repositories of every possible practice in documentation.
  • Repeating instructions in multiple locations.
  • Verbose examples that limit exploration space.

Tooling — claude doctor / /doctor

Anthropic introduced claude doctor (invoked as /doctor inside Claude Code) as the operational answer to the rightsizing problem: it automatically helps rightsize skills and CLAUDE.md files for the newer model class. First-party version of Pocock's deletion test from the Skill Checklist (Pocock).

Load-bearing convergences with the vault

This blog is the second-substantive Anthropic-official artefact on the same thesis Shihipar delivered at WF 2026 — but it moves the vault forward by naming the shifts as a formal, published set. Convergences worth flagging:

  • Field Guide to Fable (Thariq Shihipar, Anthropic) — the earlier same-author artefact. WF talk had the 80% reduction claim; this blog gives it official-documentation status and adds the five-plus-one pattern shifts.
  • Capability Overhang — the mechanism that makes constraint-removal safe. This blog is the second-source Anthropic-official framing.
  • Skill Checklist (Pocock) — the deletion test Pocock proposed at WF 2026 is now first-party-endorsed via the "remove don't add" discipline plus the claude doctor tool. Independent-vendor convergence on the same rubric.
  • Spec-Driven Development (SDD) and specifically GSD (context rot) — Christopherson's fresh-instance-per-task discipline lines up with rule #3 (progressive disclosure) and rule #4 (no repetition); both fight the same failure mode from different sides.
  • Skill Engineering — the community-consensus framing at WF 2026 gets a concrete Anthropic-side rulebook here; Tyler Brown's "revisit skills each model release" rule is the operational cadence, this blog is the specific rulebook for the current release.
  • Design Concept (Brooks) — front-load alignment, not artefact. The blog's "prefer code references over descriptions" prescription is the practical operationalisation of the design-concept-first move on the reference-material side.
  • Context Engineering — the parent discipline. This blog adds context rightsizing as a specific sub-discipline: not just "get the right context" but "remove context the newer model doesn't need." See Context Rightsizing for the vault's concept page synthesising this.

Direct bearing on this vault's own CLAUDE.md

This vault's CLAUDE.md (~370 lines, richly prescriptive) is a first-generation-era artefact by the blog's rubric — it has heavy negation phrasing ("Never modify raw/", "NEVER create documentation files"), maintains a central repository of every operation, and stores specs as markdown rather than code references. On the current Claude 5 rubric it likely earns a /doctor pass. Not a call to shorten it wholesale — the vault's CLAUDE.md also encodes non-obvious gotchas the blog explicitly says are the right thing to keep — but a natural next follow-up is a targeted rightsizing pass against the six-shift rubric. See Context Rightsizing for the vault's operational framing.

Cross-links

Source

  • Trigger: Daily Learning Telegram capture raw/processed/Daily Learning 2026-07-25 06-49 #3748.md (msg #3748, user framing: "very helpful in understanding how prompt best practices have evolved with latest models").
  • Primary: Thariq Shihipar, The new rules of context engineering for Claude 5 generation models, Claude by Anthropic blog, 2026-07-24.
  • Amplifier: Thariq on X — 5:45 PM Jul 24, 2026, 700K views.