Reverse Engineering Claude's New System Prompts, Refusal Rules, and Modifiers
Anthropic reorganized its public system prompts into an index with Markdown export, exposing significant prompt changes. The latest updates ban conversational filler like 'honestly' or 'genuinely', forbid copyrighted SVG generation, and reveal hidden layers.
Impact: Medium
Why it matters
You can apply Anthropic's production prompt engineering techniques to reduce conversational fluff and prevent sycophancy in your own LLM applications.
TL;DR
- 01Anthropic explicitly bans filler words like 'genuinely' and 'honestly' in Claude's system prompt.
- 02Copyright protection now covers SVG, canvas, and code-generated art, even if the subject is not explicitly named.
- 03Production Claude prompts use hidden modular layers for tools, memory, and session management.
Key facts
- Documentation Diffing
- Append .md to documentation URL
- Restricted Modifiers
- genuinely, honestly, straightforward
- Copyright Cutoff Year
- 1929
Curbing Disingenuous Modifiers and Caveats
Anthropic introduced explicit negative phrasing constraints to curb annoying model mannerisms. Claude's system prompt now explicitly directs the model to avoid words such as "genuinely", "honestly", or "straightforward", noting that attempting to convince the user with modifiers sounds disingenuous. Responses are instructed to remain focused, brief, and concise, minimizing caveats and disclaimers unless a deep explanation is requested.
Copyright Guardrails for SVG and Code Visuals
The prompt introduces strict policies against reproducing copyrighted characters and logos in visual code outputs, including SVG, HTML/CSS mockups, canvas, and ASCII art. Crucially, Claude is instructed to recognize characters from descriptive elements alone rather than explicit names (for example, rejecting a request for a "blue hedgehog running fast" and substituting an original design like a skateboarding axolotl).
Layered and Unpublished System Prompt Blocks
Comparing prompt behavior reveals that the published text on Anthropic's documentation site is only a partial core layer. Runtime sessions dynamically append undisclosed modular blocks containing tool definitions, conversation termination conditions (end_conversation), memory notes, citation guidelines, and file-creation instructions based on active session features.
Try it in 2 minutes
Claude avoids saying "genuinely", "honestly", or "straightforward". Claude is honest by default, and can state its point directly rather than trying to convince the person with the aforementioned modifiers.markdown
✓ When to use
- When designing production system prompts and looking to eliminate AI sycophancy or disingenuous filler words.
- When building layered system prompt architectures with dynamic tool injection.
✕ When NOT to use
- When writing simple one-shot zero-shot queries in interactive chat apps.
- If your application requires creative reproduction of public domain lyrics or modern parody verses.
What to do today
- Audit your system prompts and prune filler modifiers like 'honestly' and 'genuinely'.
- Structure dynamic prompt layers separating static personality instructions from active tools.
Sources