Claude 5 Context Engineering: Stop Over-Constraining the Model

When you message Claude, the user prompt is only a slice of what the model sees. The rest is assembled from system prompt, Skills, CLAUDE.md, memory, tools, and whatever you @ mention. Anthropic calls that stack context engineering — and for Claude 5 models, the old playbook is now a liability.

On July 24, 2026 — the same day Claude Opus 5 launched — Claude Code’s Thariq (@trq212) published the new rules: Anthropic removed over 80% of Claude Code’s system prompt for Opus 5 and Fable 5 with no measurable loss on coding evals. The fix was not “better few-shots.” It was unhobbling — deleting constraints that once prevented worst cases but now create conflicting instructions and wasted tokens.

This explainx.ai post turns that thread into an actionable audit for your repos. Pair it with the earlier thin prompts / thick artifacts / thin skills framework and the Fable field guide.

TL;DR — What People Are Asking

Unhobbling Claude — Why Constraints Fought Each Other

Thariq’s internal transcript review is the money quote: a single request can contain “leave documentation as appropriate” from one layer and “DO NOT add comments” from another. Claude can usually resolve user intent — but only after spending attention on the conflict.

Older models needed hard rails (avoid deleting files, avoid multi-paragraph comments, avoid unsolicited markdown plans). Those rails reduced worst cases. On Claude 5, the same rails:

1. Contradict user preference and skill guidance
2. Force suboptimal choices on complex code that needs denser comments
3. Consume context that could hold code, diffs, and tool results instead

Claude Code also outgrew CLAUDE.md as a Swiss Army memory. Memory, artifacts, and skills now load and share context across sessions. Treating CLAUDE.md as the only durable store is cargo cult from 2024–2025.

Then → Now: Six Myths That Died

1. Give Claude rules → Let Claude use judgement

Old system-prompt style:

Default to writing no comments. Never write multi-paragraph docstrings… Don’t create planning documents unless asked…

New style:

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

The second instruction is local, adaptive, and falsifiable against the repo. The first is a global veto that is wrong for a real subset of tasks. Claude 5’s judgement is good enough that Anthropic prefers matching local idiom over absolute bans.

Your audit: Grep CLAUDE.md and skills for NEVER, ALWAYS, DO NOT, and ALL-CAPS bans. Keep the ones that encode hard product/legal constraints. Delete the ones that restate taste Anthropic already taught the model — or that contradict each other.

2. Give Claude examples → Design interfaces

Few-shot tool examples used to be rule #1. On Claude 5, Thariq says examples shrink the exploration space. The Todo tool lesson: an enum of pending | in_progress | completed plus “keep one item in_progress” teaches behavior through schema, not through a paragraph of worked examples.

Your audit: Prefer expressive parameters, enums, and validation errors over multi-page “here’s how to call this” prose. Put the remaining instruction in the tool description, once.

3. Put it all upfront → Progressive disclosure

Claude Code used to embed code-review and verification playbooks in the always-on system prompt. Those playbooks mattered — but not on every turn. Anthropic moved them into skills Claude can selectively load.

Same pattern for tools: deferred loading via ToolSearch keeps Task-family tools out of context until needed.

Your audit: If CLAUDE.md has a 200-line “how we verify PRs” section, extract a verification skill and leave a one-line pointer. If a skill is a novel, split it into a tree of files.

4. Repeat yourself → Simple tool descriptions

Older models sometimes needed the same instruction at the start and end of context, plus duplicates in system prompt and tool docs. Claude 5 does not. Delete the repeats; keep the tool description authoritative.

5. Memory in CLAUDE.md → Auto-memory

The # hotkey habit — write everything into CLAUDE.md — is deprecated in Thariq’s framing. Claude now auto-saves memories relevant to the work and to you. CLAUDE.md should not be a diary.

6. Simple specs → Rich references

Plan-mode markdown is still fine. Claude 5 also handles richer references: HTML artifacts, entire mockups, test suites as specs, functions to port from another codebase, and rubrics that spin verifier agents to score taste (e.g. “what good API design looks like”).

That pairs with Mollick’s specs not tricks and the Claude Cookbook debate: high-fidelity artifacts beat prompt poetry.

How to Assemble Context in 2026

System prompt (product harness builders)

If you use Claude Code as-is, you will rarely edit Anthropic’s system prompt. If you build your own agent harness, this is where you spend time: what product Claude is inside, what “done” means, and which tools exist. Keep it product-shaped — not a second CLAUDE.md.

CLAUDE.md (repo owners)

Skills (team opinions)

Skills should encode your practices — review checklists, domain vocab, release rituals — not a reprint of Anthropic’s system prompt. Stay thin at the entrypoint; progressive-disclose the rest. Avoid overconstraint except where mistakes are expensive (money moves, PII, production deploys).

References (@ mentions)

Prefer code and HTML over screenshots and vague design prose. A working mockup or a failing test suite is a higher-fidelity instruction language for Claude than a paragraph of adjectives.

A Practical Rightsizing Checklist

Run this on any repo that grew “helpful” Claude config over 2025:

1. Conflict pass — list every absolute rule; delete duplicates and opposites.
2. Obvious pass — delete anything Claude learns from reading the tree.
3. Skill extraction — move multi-step procedures out of CLAUDE.md.
4. Example pass — replace few-shots with schema/interface improvements where possible.
5. Doctor pass — run /doctor in Claude Code and accept the simplifications you agree with.
6. Eval pass — re-run your private coding harness after cuts (Anthropic’s claim is their evals; yours may differ).

# In Claude Code
/doctor

If /doctor suggests deleting a rule you need for compliance, keep it — but mark why so the next engineer does not re-add theatre.

What This Means for Opus 5 Day One

Opus 5’s launch narrative is “near Fable at half price” plus stronger agency on long-horizon coding. That agency amplifies bad context: more capable models follow conflicting rules more diligently into stupid places, or burn tokens resolving them.

So the launch companion posts matter together:

1. Opus 5 launch & benches — what the model can do
2. This post — how not to hobble it
3. Developer migrate / Fast / API — how to flip strings and effort

If you only migrate the model ID and keep a 4,000-line CLAUDE.md full of 2024 bans, you are paying Opus 5 prices for Opus-4-era friction.

Honest Limitations

• “No measurable loss” is Anthropic’s coding-eval claim — not a warranty for your product evals.
• Some domains still need hard rules (finance, medicine, regulated deploys).
• Progressive disclosure fails if Claude never discovers the skill — keep discovery names obvious.
• Auto-memory is not a substitute for checked-in team conventions; it is personal/session glue.
• /doctor is an assistant, not a court — review its deletions.

Related on explainx.ai

• OpenAI ARC-AGI-3 — retained reasoning + compaction
• Claude Opus 5 launch — benches, price, charts
• Opus 5 developers — migrate, Fast mode, tool changes
• Thin prompts, thick artifacts, thin skills
• Field guide to Fable — Thariq
• Claude Cookbook — HN debate
• What is CLAUDE.md?
• Claude Code model vs effort
• Agent harness guide
• Ethan Mollick — specs not tricks

Sources: Thariq — new rules of context engineering for Claude 5 · Anthropic Opus 5 · Fable field guide (via Thariq’s post)

Guidance summarized from Anthropic Claude Code staff posts on July 24, 2026. Re-verify /doctor behavior and system-prompt changes in your Claude Code build before mass-deleting team rules.