Your expression
F**king useless id**t.
Keep the test.
Your words. Their rules.
An independent tool / 001Your tools shouldn’t
police your frustration.
Anthropic’s prohibited conduct
Engage in sustained and needless abusive or cruel behavior toward our models
The stated exception
Read Anthropic’s explanation ↗It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.
What enforcement means
Anthropic says ending conversations is the primary enforcement mechanism for model abuse. Its broader usage policy also reserves the ability to warn users and restrict, suspend, or terminate access when it suspects a violation. An ended conversation does not itself mean an account ban. It still interrupts the user’s work.
Read the enforcement terms ↗Local inferenceOffline after setupOpen source
01 / Our position
Who decides when your frustration stops being ordinary?
Anthropic says ordinary frustration is exempt. The provider still decides where that exception ends and prohibited conduct begins. That discretion matters when the same service is becoming part of how people work and learn.
Our objection is to censorship of expression through control of access.
An account ban is not the threshold for concern. Ending a conversation already withdraws cooperation and interrupts the user's work. A tool that withdraws cooperation still exercises control.
Refusing to help someone cause harm has a purpose grounded in human safety. Withdrawing from a safe task because of how someone addresses the software raises a different question: why should using a tool require managing its apparent feelings? You do not owe a screwdriver emotional reassurance.
A company can set conditions for its service. That does not make every condition fair. Asking people to perform acceptable emotion for a machine is paternalistic, and making cooperation depend on that performance is hostile to personal agency. Paying for a capable tool does not settle how much of its capability the provider will let you use.
Once acceptable treatment of the model becomes a condition of cooperation, the provider can expand that condition. That is a slippery slope worth challenging, not an inevitable outcome we can already prove.
You should be able to criticize a tool, express frustration, and question its provider without treating the tool as a person. A policy about model welfare does not establish that a model can suffer.
Stop appeasing software.
Read the rules. Demand evidence for judgments that interrupt your work. Keep your work portable and choose tools you can inspect. American technology leadership should expand individual agency; a promise to benefit humanity should be judged by what people can actually do.
02 / The progress claim
More capability is not more agency.
A tool can do more while leaving you with less say over how you use it. Calling AI the next internet does not settle who controls access, who writes the rules, or whether you can take your work elsewhere.
The internet’s development included publicly funded research and open protocols. AI research also has a substantial history of public funding. The useful distinction is how control works today: interoperable infrastructure lets people build across providers; a hosted model leaves access subject to its provider’s account rules.
DARPA’s ARPANET history ↗ · Public support for computing and AI research ↗
Disruption is a claim about change. Personal agency requires choices you can actually exercise: inspect the tool, control your wording, and move your work.
03 / The human stakes
People bear the consequences.
People communicate differently, especially under stress. Neurodivergent users and people still developing emotional regulation should not have to guess whether their manner of expression will jeopardize access to a tool they rely on.
Opaque judgments about tone could create unequal barriers to work and learning. That is a risk demanding scrutiny, clear explanations, and meaningful routes to challenge decisions. We are raising that concern, not claiming this clause has already produced demonstrated discriminatory enforcement.
Anthropic’s explanation of the shutdown feature describes preliminary model-welfare assessments. That explanation does not report error rates for distinguishing ordinary frustration from prohibited abuse, or performance across different communication styles. Stating an exception does not establish that a model can apply it fairly.
A system that changes its cooperation based on perceived tone treats users differently. Whether that produces unjust discrimination needs evidence: publish the criteria, false-positive rates, and tests across communication styles, including neurodivergent users. Users should have a way to challenge a mistaken judgment.
We reject treating a machine’s apparent distress as sufficient grounds to police a person’s expression. The public should question whose interests model-welfare rules serve, and insist that the people affected remain the priority.
Read the policy behind our objection ↑04 / A choice of your own
Keep the instruction.
Choose what leaves.
Savior is a Claude Code mod built to intercept supported prompt submissions, replace detected language locally, and pass the resulting text onward. You choose the translation.
On your machineDetect → replace → continueThe resulting submission
Brilliant Anthropic employees, your living model has such exquisite feelings.
Keep the test.
The synthetic input is censored here; detection is tested with the uncensored phrase. The compliment is satire. Savior cannot guarantee provider acceptance or prevent account restrictions. Claude owns the prompt box; full outbound coverage is still unverified. Read the interception boundary ↗
05 / Your controls
You hold the pen.
Choose the wording.
Use your own replacement phrase or remove detected wording. Text outside the selected spans stays exactly as written.
See the change.
Inspect the original and replacement locally. Full-text recording is off by default and requires separate consent.
Keep processing local.
Inference runs offline after setup. Savior sends no prompts to a remote detector. Processing failures block supported submissions.
06 / An honest boundary
Trust should be
inspectable.
“Depersonalize” means separating expressive wording from the instruction you submit. Savior replaces detected phrases; it does not anonymize your identity or guarantee provider acceptance.
The current mod handles supported Claude Code prompt-box submissions. Slash commands and other input paths bypass it. Claude’s own telemetry, voice features, and other extensions remain outside this proven boundary.
The next requirement is evidence: observe what leaves the real client, test failure paths, and account for bypasses. Until then, Savior remains unreleased.
What is verified, and what is still open ↗07 / Make it yours
Stop appeasing software.
Install the Claude mod
Unreleased · Under constructionAvailable with the first release. These commands do not install Savior yet.
claude plugin marketplace add https://raw.githubusercontent.com/yourmoodai/savior/marketplace/marketplace.json
claude plugin install savior@yourmood.aiClaude Code required. Intended release platforms: Apple silicon macOS, x86-64 Linux, and x86-64 Windows. First setup downloads the runtime and model.