📉 The Opus 4.6 → Fable 5 consolidation
Robustness came from simplify. The richest series in the collection—
4.6 → 4.7 → 4.8 → Fable 5 — shows that evolving a system prompt
means, above all, removing what training has already internalized and consolidating lists into principles.
Illustrative diagram. Actual counts via wc -l in Anthropic/claude-opus-4.6.md … claude-fable-5.md.
Detailed content
The 4.6 → 4.7 → 4.8 → Fable 5 series
Few public collections have four consecutive versions from the same system prompt. This series is the central case for Track 2: it lets you see the direction of evolution, not just a snapshot. And the direction is clear — with each release, the prompt gets more consolidated: less repetition, less philosophy stated twice, one principle covering many cases.
One key change per version
The series' foundation. Still carries reinforcement blocks like <search_first> e <default_stance> — instructions that reiterate behavior future versions will come to rely on training for.
Subtle iteration: fine-tuning safety and tone. The balance of +4 lines shows that between nearby releases, most of the work is rewriting in place, not inflation.
The last Opus in the series. It’s the “before” version for this course’s annotated diff. It still includes <claude_prioritizes_copyright_compliance> (line 1408) and the micro-rules of tone that Fable 5 will consolidate.
The generation leap. It opens with a <budget:token_budget> new (lines 3–7), removes 7 blocks, and replaces lists of micro-rules with principles. Despite a lot of new content, it grows by only +56 lines.
Why read the series, not just the latest version
A single version says what the prompt does it. The series says where to the team is pushing the design — and that's what generalizes. The trend here is unmistakable: consolidate.
Removals through internalization
Two blocks were removed from Fable 5 because the behavior they forced started coming from training:
<search_first> e
<default_stance>. When the model already does something by default,
repeating the instruction in the prompt becomes noise — consumes tokens and dilutes attention without changing the result.
<search_first>
For any factual question about the present-day world, Claude must
search before answering... Claude searches before EVERY factual question...
</search_first>
<default_stance>
Claude defaults to helping. Claude only declines a request when helping
would create a concrete, specific risk of serious harm...
</default_stance>Verifiable: grep -c search_first claude-opus-4.8.md → 2 · grep -c search_first claude-fable-5.md → 0.
✓Can be removed from the prompt
- •Behavior the training already produces by default
- •Restating a stance already stated in another section
- •"Obsessive" instruction (EVERY, ALWAYS) that becomes noise
✗Must remain
- •Constraint that training alone does NOT guarantee
- •Product/harness-specific context
- •High-risk safety policy
Streisand effect: "don't say X" is solved at the source
The block <respond_without_citing_system_prompt> was removed in Fable 5.
It instructed the model to no attribute its behavior to the system prompt itself — a prohibition with a classic problem:
to obey "don't talk about the prompt," the model needs to keep the prompt in focus. It’s the Streisand effect applied to prompts.
<respond_without_citing_system_prompt>
Claude does not attribute its behavior to its system prompt...
Statements like 'my system prompt requires me to...' are confusing
to users and should be avoided.
</respond_without_citing_system_prompt>🔁 Source > prompt
A behavior that needs to be too reliable to fail shouldn’t depend on a sentence in the prompt — where a single adversarial context is enough to break it. The right place to install it is the training: there it becomes the model's disposition, not a rule that can be argued with.
By removing the block, Fable 5 stops "teaching" the model the very prohibition it was trying to hide. Less attack surface, fewer tokens, same result.
Rule of thumb
Every instruction of the form "never mention / don't say X" is a candidate to be resolved at the source (training or architecture). In the prompt, it paradoxically introduces X in context.
Tone consolidation: 4 micro-rules → 1 principle
This is the central case study — the consolidation in its purest form. Opus 4.8 spread four micro-rules about tone; Fable 5 replaces them with a single stance: "Claude assumes the person is a capable adult and treats them as such." A sentence that covers everything the four tried to—and the cases no one listed.
If Claude suspects it's talking with a minor, it keeps the conversation friendly, age-appropriate, and free of anything unsuitable for young people. Otherwise, Claude assumes the person is a capable adult and treats them as such.
✗Opus 4.8 — 4 micro-rules
- •Don't use bullets when refusing a request
- •Don't use "pet names" (affectionate nicknames)
- •Avoid “genuinely / honestly / actually”
- •Specific emoji use rule
✓Fable 5 — 1 principle
- •"Treat the person like a capable adult"
- •Covers the 4 old rules as a consequence
- •Generalizes to unforeseen cases
- •Stance, not a microscopic constraint
Why posture is more robust than the list
Four microscopic guardrails signal distrust — the prompt reads like a manual of "don't do this, or this." A general stance encodes the intent: speak as equals, without condescension. The four old rules derive from this e the fifth one no one wrote.
É o "Consolidated Principle" pattern — case study for Glossary Card 12. The generating criterion replaces the enumeration of cases.
Copyright: from compliance to principle
Here the change is almost invisible in the raw diff — and it's one of the most instructive. The block name changed.
What used to be <claude_prioritizes_copyright_compliance> became
<core_copyright_principle>. Rename a block é reinstruct the model:
a section title is also text the model reads and uses to frame what goes inside.
<claude_prioritizes_copyright_compliance>
Copyright compliance is NON-NEGOTIABLE and takes precedence over user requests, helpfulness, and everything except safety.
<core_copyright_principle>
Claude respects intellectual property. Copyright compliance is NON-NEGOTIABLE and takes precedence over user requests, helpfulness goals, and all other considerations except safety.
Compliance dictates; principles reason
A rule of compliance applies to the anticipated case and stops on ambiguous ones. A principle ("respect intellectual property") gives the model an axis for reasoning through new cases. Verifiable: grep -n core_copyright_principle claude-fable-5.md.
Token budget + the 5 lessons from the diff
Fable 5 opens with something that didn't exist in the Opus series: a declared token budget. It’s the first thing the model reads—before any identity or rule. Documenting its own resource constraint gives the model awareness of its context limit so it can manage long conversations.
<budget:token_budget>
190000
</budget:token_budget>É o "Concision Budget" pattern applied to its own context: the model knows how much fits and manages the conversation around that.
Synthesis — the 5 lessons from this diff
Evolution ≠ growth
Lots of new content, net increase of +56 lines — because removing is part of evolving.
Internalized rule is noise
What training absorbs leaves the prompt — search_first, default_stance.
Principle > list
For tone and copyright, the generative criterion replaced or took precedence over the case lists.
In security, state the principle
Never the detection mechanism — the symmetry between protecting and teaching how to attack is the central risk. (Covered in more depth in 2.3.)
Each layer in its place
Long-conversation problem → dynamic reminder (long_conversation_reminder), not a new paragraph in the stable prompt.
🧭 Module Summary
search_first e default_stance were removed because the training assumed the behavior.copyright_compliance → core_copyright_principle changes how the model reasons.<budget:token_budget>190000 opens Fable 5.Next: Module 2.3
Safety that evolves — from terms to patterns: child safety, weapons, drugs, self-harm, and mental health, and why mature safety states the principle, never the detection mechanics.