Claude Opus 5 Feels Shallow? Fix Legacy Rules (2026)

Common symptoms after switching to Opus 5

After switching Claude Code to Claude Opus 5, teams often report symptoms such as:

📑Table of Contents

1. Why Opus 5 can feel shallow — lean prompt vs legacy rules
2. Official Opus 5 / Fable 5 prompting principles
3. Comparison table — legacy harness vs Opus 5 lean harness
4. CLAUDE.md rewrite path — remove, add, then rightsize
5. Adoption checklist — try, fix, or roll back
6. FAQ
7. Takeaways

• answers feel “shallow”
• the agent acts fast but verifies little
• narration and tokens balloon without better outcomes

---

First hypothesis and article scope

That can look like model regression. A more reproducible first hypothesis is a harness mismatch: Claude Code’s lean system prompt default colliding with long CLAUDE.md / rules written for the Opus 4.x era.

This article pins the facts with the Claude Code changelog, Anthropic’s Prompting Claude Opus 5 and Fable 5 docs, and independent coverage from ITmedia AI+, THE DECODER, AI Revolution, and operator notes (masa_wunder). It then turns those sources into remove/add guidance and a one-task measurement checklist.

---

Learning outcomes

---

Why Opus 5 can feel shallow — lean prompt vs legacy rules

The short version: loading Opus 4.7-era behavioral scaffolding onto Fable 5 / Opus 5 lean defaults changes agent behavior. The model is not necessarily “dumber.” More often, the product system prompt got thinner while user rules stayed thick.

Models where lean system prompt is default

Claude Code’s public changelog states:

The lean system prompt is now the default for all models except Haiku, Sonnet, and Opus 4.7 and earlier

Operational takeaways:

• Source: Claude Code changelog
• For models outside that exclusion set, lean is default
• That is a concrete reason teams see different rules behavior on Opus 5 than on Opus 4.7

---

Independent reporting on the ~80% cut

Independent reporting points the same way.

• ITmedia AI+ (2026-07-09) covered Anthropic’s Tariq Shihipar remarks: Claude Code cut roughly 80% of its system prompt and stopped supplying behavioral examples that over-constrained newer models.
• THE DECODER (2026-07-02) reported the same ~80% reduction, noted that Fable 5-class models prefer smaller system prompts, and described a shift from hard “do not” rules toward context steering.
• AI Revolution ties the shipped changelog language (v2.1.154 lean default) to Fable 5-era design and warns that over-prescriptive prior-model Skills / CLAUDE.md can hurt quality.

---

“No prompts needed” is the wrong takeaway

“No prompts needed” is the wrong takeaway. After the product prompt lean-down, keeping CLAUDE.md clauses such as the following tends to create two failure modes:

• mandatory final verification
• always-on fresh subagent review
• long ban lists plus example dumps

The two common outcomes:

1. Over-constraint: fights default self-checks and act-first behavior, inflating tokens and narration
2. Under-steering on intent: outputs look abrupt or shallow because the harness still teaches the old control style

Before labeling a capability drop, inspect the harness.

---

Official Opus 5 / Fable 5 prompting principles

Prompting Claude Opus 5

Anthropic’s Prompting Claude Opus 5 asks operators to adjust scaffolding.

---

Brevity and verification

• Brevity needs an explicit ask: default user-facing answers can run long; effort controls thinking volume, not display length
• Remove over-verification: Opus 5 already self-checks; standing “final verification,” “double-check,” or “fresh subagent review” steps waste tokens

---

Scope, delegation, and cadence

• State scope: on narrow tasks, forbid unsolicited expansion
• Limit subagent delegation: reserve subagents for large independent parallel work; do not delegate few-tool chores or self-verification
• Set progress cadence: one line before the first tool, updates only when it matters, outcomes first at the end

---

Prompting Claude Fable 5

Prompting Claude Fable 5 documents complementary steering:

• act-first: “When you have enough information to act, act.”
• recommendation-not-survey: prefer a recommendation over an exhaustive option survey
• refactor over-detailed Skills written for prior models
• audit instructions that demand full chain-of-thought replay (reasoning_extraction risk)
• when capability rises, re-evaluate instructions, tools, and guardrails; add brief reason/intent, not only the request

---

Operator notes (masa_wunder)

Operator write-ups such as note (masa_wunder) translate the same official guidance into harness edits:

• drop forced verification and standing subagent review
• add brevity, delegation caps, narration grain, and scope fences

Source

They also recommend practical effort settings:

• review effort per task
• prefer low effort with thinking enabled over thinking disabled
• watch invalid combinations such as xhigh/max with thinking disabled

---

Comparison table — legacy harness vs Opus 5 lean harness

Use this table when auditing CLAUDE.md.

Legacy vs lean: system prompt, verification, length

Dimension	Legacy rules still common after 4.x	Opus 5 / lean-oriented shape
System prompt context	Long bans, procedures, example dumps	Assume lean default (4.7-and-earlier excluded); keep instructions thin
Verification	Mandatory final check / subagent review every run	Respect default self-checks; delete standing over-verification
Length and tone	Implicit long answers and heavy narration	Explicit brevity / tone preference

---

Legacy vs lean: delegation, Skills, evaluation

Dimension	Legacy rules still common after 4.x	Opus 5 / lean-oriented shape
Delegation	Subagents for small chores	Large independent parallel work only
Skills / CLAUDE.md	Over-prescriptive prior-model text	Brief instruction + intent; /doctor rightsizing
Evaluation	“Did it obey the checklist?”	Before/after lines of narration, subagent starts, token delta

---

Sources (as of July 2026)

• Claude Code changelog
• Prompting Claude Opus 5
• ITmedia AI+
• note masa_wunder
• THE DECODER
• AI Revolution

---

How to read the table

If your file still reads like the left column, Opus 5 is more likely to feel both shallow and expensive. The right column is the shared direction of official docs and independent operator notes.

---

CLAUDE.md rewrite path — remove, add, then rightsize

You do not need a full rewrite on day one. Safer order: comment out collision sources → add short steering → measure one task.

Remove candidates (collision sources)

1. Standing “always verify,” “double-check,” or mandatory final verification steps
2. Always-on “fresh subagent review” requirements
3. Long ban lists and behavioral example dumps that duplicate a lean product prompt
4. “Replay your full reasoning in the answer” style instructions (Fable 5 reasoning_extraction risk)

---

Add candidates (official-aligned steering)

1. Explicit brevity / length preference
2. Progress cadence (when to narrate one line)
3. Task scope fences (do not expand)
4. Subagent delegation limits (large independent parallel work only)
5. Short reason/intent behind the request

---

Rightsizing and ops notes

As AI Revolution summarizes, Claude Code /doctor can propose CLAUDE.md cuts for content the codebase already implies. Dropping redundant bans and restating the obvious reduces instruction density.

Keep effort per-task rather than pinned at maximum. On failure, do not throw away the entire rules file—roll back only the colliding clauses so you can isolate cause.

Treat Skills like CLAUDE.md. Fable 5 docs warn that prior-model over-prescriptive Skills can degrade output and should be refactored toward brief instruction, trigger conditions, and intent.

---

Adoption checklist — try, fix, or roll back

Try this week’s one-task experiment if most of these are true:

• You run Opus 5 or another lean-default model
• CLAUDE.md / project rules are still long 4.x-era text
• You can reproduce shallow answers, thin verification, heavy verification cost, or token inflation

Checklist

1. ✅ Confirm on the changelog whether your model uses lean system prompt by default
2. ✅ Temporarily comment out forced verification and always-on subagent review in CLAUDE.md
3. ✅ Add short brevity, scope, delegation, and cadence lines
4. ✅ On the same task, record before/after narration lines, subagent starts, and tokens
5. ✅ Run /doctor or a diff review to remove codebase-obvious prose
6. ✅ On failure, restore only the colliding clauses—not the whole file

---

When to keep a heavier path

If regulation or audit requires external verification evidence, do not outsource that gate to model defaults alone.

Split standing model instructions from a separate review lane (PR checks, human checklists, audit logs). Official guidance targets standing over-verification inside the agent prompt, not the existence of organizational quality gates.

---

Next action this week

Finish one familiar representative task with three removals + three additions + one metric set (narration lines, subagent starts, tokens). If the delta is clear, promote the same policy into the shared CLAUDE.md template.

---

FAQ

Q1. Did Opus 5 actually get “shallower”?

Official docs and independent reporting point first to a mismatch between Claude Code’s lean system prompt default and legacy over-constraining rules—not to a simple capability label. Inspect the harness first.

---

Q2. Is lean system prompt on for every model?

The changelog says lean is default for all models except Haiku, Sonnet, and Opus 4.7 and earlier. Use that exclusion when explaining 4.7 vs 5 behavior differences.

---

Q3. Can we delete all verification steps?

Opus 5 docs warn that standing final verification / subagent verify causes over-verification. Keep audit-required review in a separate process; do not confuse it with always-on model instructions.

---

Q4. Can Skills stay unchanged?

Fable 5 docs say prior-model over-prescriptive Skills can degrade quality and should be refactored. AI Revolution makes the same caution. Compress toward brief instruction plus intent.

---

Q5. How do we measure success?

As operator notes describe, compare the same task on narration line count, subagent starts, and token delta. Cost and observability beats a pure “did it obey the old checklist?” score.

---

Takeaways

• The leading hypothesis for “shallow Opus 5” is lean system prompt × legacy rules collision
• Follow official Opus 5 / Fable 5 guidance: cut forced verification and over-detailed Skills; add brevity, scope, delegation limits, cadence, and intent
• Lock facts with changelog + independent reporting, then finish one measured task rewrite
• Keep audit verification in a separate lane from standing model instructions

For adjacent harness topics, see CLAUDE.md design patterns, Claude Opus 4.7 and /effort, and writing Claude Code Skills.

Related articles:

• 13580
• 13621
• 13549