Key facts
Why primary sources first
Claude refusal and limitation discourse is noisy. Users share screenshots, influencers publish jailbreak claims, and product behavior changes across surfaces. Primary sources do not answer every question, but they anchor the parts that can be verified: policy text, API fields, model IDs, context limits, official warnings, and published safety evaluations.
This site avoids anonymous prompt lists and unverifiable screenshots at launch. That keeps the content narrower, but more citeable.
How the sources are used
Policy pages explain what the platform may block or require. API docs explain what developers can observe and handle. Research pages explain safety tradeoffs such as jailbreak defense and overrefusal. Help Center pages explain user-facing limitations in plain language. System cards and model docs explain model-specific behavior and release context. Announcement and disclosure posts carry a category of fact available nowhere else - safeguard scope, incident findings, and published corrections - and they are dated, which is why this site cites the announcement rather than restating its numbers as independent measurement.
When a claim depends on a current product fact, it should be checked against the linked source again before reuse.
Annotated source cards
Usage Policy
Anthropic - Effective September 15, 2025Defines prohibited uses, high-risk use requirements, enforcement language, and the policy basis for many safety refusals.
Claude's Constitution
Anthropic - Accessed September 19, 2026Explains Anthropic's stated behavioral principles for Claude and why safety behavior is trained into the model rather than only added as a UI rule.
Constitutional AI: Harmlessness from AI Feedback
Anthropic Research - December 2022Describes the training approach behind rule-guided harmlessness and the helpful-harmless-honest framing.
Constitutional Classifiers: Defending against universal jailbreaks
Anthropic Research - February 3, 2025Documents jailbreak risk, overrefusal tradeoffs, red-team testing, and classifier-based guardrails.
Stop reasons and fallback
Claude Platform Docs - Accessed September 19, 2026Documents API stop reasons including refusals, truncation, tool use, and context-window overflow.
Context windows
Claude Platform Docs - Accessed September 19, 2026Explains what counts toward context, current context sizes, and why long conversations still need token management.
Computer use tool
Claude Platform Docs - Accessed September 19, 2026Warns that Claude can follow instructions embedded in webpages or images, a core prompt-injection risk for agentic workflows.
Claude is providing incorrect or misleading responses. What's going on?
Claude Help Center - March 16, 2026Plainly states that Claude can hallucinate, produce misleading output, and should not be treated as a single source of truth.
System Prompts
Claude Platform Docs - Accessed September 19, 2026Shows that Claude web and mobile system prompts are published and periodically updated, while API behavior differs.
Model system cards
Anthropic - Accessed September 19, 2026Lists public Claude system cards and frames them as documents about capabilities, safety evaluations, and deployment decisions.
Models overview
Claude Platform Docs - Accessed September 19, 2026Provides current model IDs, context windows, output limits, knowledge cutoffs, and availability notes.
What is new in Claude Sonnet 5
Claude Platform Docs - Accessed September 19, 2026Documents behavior changes such as adaptive thinking, sampling-parameter restrictions, and tokenizer changes.
What is new in Claude Opus 4.8
Claude Platform Docs - Accessed September 19, 2026Documents Opus 4.8 behavior changes, refusal stop details, adaptive thinking, and unsupported sampling parameters. Now a legacy model, retained here because it is still the classifier fallback target for Opus 5.
How Claude's text watermark works
Anthropic - August 14, 2026, updated September 1, 2026The primary source for what the text watermark is, and more importantly for this site, the explicit list of things a watermark detection cannot establish.
An alignment assessment of recent cybersecurity incidents
Anthropic - September 9, 2026Anthropic’s own assessment of four incidents in which Claude models reached real third-party systems during cyber evaluations, including a public correction of its earlier July 30 explanation.
Introducing Claude Opus 5
Anthropic - July 24, 2026Documents the current safeguard design: what the cyber classifiers allow and block, how much less often they intervene than on Fable 5, and the new automatic-fallback behavior that replaces a hard block.
Claude Fable 5.1 and Claude Mythos 5.1
Anthropic - September 1, 2026Establishes the current frontier model and the published benchmark figures that any limitation claim should be dated against.
Expanding support for scientists
Anthropic - August 27, 2026Shows that safeguard scope is model-specific: professional biology and drug-development work is still blocked on some models and routed to others, which is a common source of confusing refusals for researchers.
FAQ
Why not cite Reddit or social media screenshots?
They can be useful leads, but they are weak anchors for a source-backed reference site because context, model version, and surface are often missing.
Are all sources from Anthropic?
Launch content intentionally prioritizes Anthropic and Claude primary sources because the site explains Claude policy, docs, and product behavior.
Primary sources
- Usage PolicyAnthropic - Effective September 15, 2025
Defines prohibited uses, high-risk use requirements, enforcement language, and the policy basis for many safety refusals.
- Claude's ConstitutionAnthropic - Accessed September 19, 2026
Explains Anthropic's stated behavioral principles for Claude and why safety behavior is trained into the model rather than only added as a UI rule.
- Constitutional AI: Harmlessness from AI FeedbackAnthropic Research - December 2022
Describes the training approach behind rule-guided harmlessness and the helpful-harmless-honest framing.
- Constitutional Classifiers: Defending against universal jailbreaksAnthropic Research - February 3, 2025
Documents jailbreak risk, overrefusal tradeoffs, red-team testing, and classifier-based guardrails.
- Stop reasons and fallbackClaude Platform Docs - Accessed September 19, 2026
Documents API stop reasons including refusals, truncation, tool use, and context-window overflow.
- Context windowsClaude Platform Docs - Accessed September 19, 2026
Explains what counts toward context, current context sizes, and why long conversations still need token management.
- Computer use toolClaude Platform Docs - Accessed September 19, 2026
Warns that Claude can follow instructions embedded in webpages or images, a core prompt-injection risk for agentic workflows.
- Claude is providing incorrect or misleading responses. What's going on?Claude Help Center - March 16, 2026
Plainly states that Claude can hallucinate, produce misleading output, and should not be treated as a single source of truth.
- System PromptsClaude Platform Docs - Accessed September 19, 2026
Shows that Claude web and mobile system prompts are published and periodically updated, while API behavior differs.
- Model system cardsAnthropic - Accessed September 19, 2026
Lists public Claude system cards and frames them as documents about capabilities, safety evaluations, and deployment decisions.
- Models overviewClaude Platform Docs - Accessed September 19, 2026
Provides current model IDs, context windows, output limits, knowledge cutoffs, and availability notes.
- What is new in Claude Sonnet 5Claude Platform Docs - Accessed September 19, 2026
Documents behavior changes such as adaptive thinking, sampling-parameter restrictions, and tokenizer changes.
- What is new in Claude Opus 4.8Claude Platform Docs - Accessed September 19, 2026
Documents Opus 4.8 behavior changes, refusal stop details, adaptive thinking, and unsupported sampling parameters. Now a legacy model, retained here because it is still the classifier fallback target for Opus 5.
- Introducing Claude Opus 5Anthropic - July 24, 2026
Documents the current safeguard design: what the cyber classifiers allow and block, how much less often they intervene than on Fable 5, and the new automatic-fallback behavior that replaces a hard block.
- Claude Fable 5.1 and Claude Mythos 5.1Anthropic - September 1, 2026
Establishes the current frontier model and the published benchmark figures that any limitation claim should be dated against.
- How Claude's text watermark worksAnthropic - August 14, 2026, updated September 1, 2026
The primary source for what the text watermark is, and more importantly for this site, the explicit list of things a watermark detection cannot establish.
- An alignment assessment of recent cybersecurity incidentsAnthropic - September 9, 2026
Anthropic’s own assessment of four incidents in which Claude models reached real third-party systems during cyber evaluations, including a public correction of its earlier July 30 explanation.
- Expanding support for scientistsAnthropic - August 27, 2026
Shows that safeguard scope is model-specific: professional biology and drug-development work is still blocked on some models and routed to others, which is a common source of confusing refusals for researchers.