You choose when prompt text leaves your device.
The CasePrompts public tools are built local-first: the work happens in your browser, and prompt text reaches a cloud model only when you choose an optional AI feature. CasePrompts does not intentionally persist those requests or results. Provider and hosting-infrastructure logs may still be retained under the configured services' account terms. Here's exactly what that means.
CasePrompts implementation review: . Provider-source review: . Provider terms can change; confirm the current account and documentation before entering sensitive information.
01
The builder runs in your browser
Assignment details you choose to enter—prefer aliases, issue labels, and only the source context the assignment needs—are assembled on your device. The free builder does not upload prompt-field contents or send them to a CasePrompts API. Your browser may remember safe preferences, such as your last-used practice, under its own local-storage settings.
Your inputs
aliases · issue labels · source context
CITE assembly
on your device
Finished prompt
ready to copy
Prompt-field contents are not uploaded by the free builder.
Ask your files: the file never leaves the tab
Preview
When you open a transcript or record in Ask your files, your browser reads it into memory in this tab. Text extraction, page-and-line detection, and search all run on your device. The file and your questions are not uploaded, not written to browser storage, and not placed in a URL; closing or reloading the tab clears them.
- Find passages works on every device and makes no network request with your file.
- Copy cited prompt puts only the passages you select on your clipboard, for an AI tool your firm has approved. What happens next is governed by that tool.
- On-device AI is off until you turn it on. It downloads an open model once from Hugging Face and GitHub (they see your IP address, not your files or questions) and runs it in an isolated worker whose network access is limited, in code and by the browser's content security policy, to downloading that model. You can delete the download at any time.
Analytics, where enabled, record only outcomes such as “files loaded” or “search returned results”—never file names, sizes, page counts, questions, or text.
02
Prompt text reaches the AI route only when you choose
The free builder and library assemble prompts in your browser. The only features that send prompt text to a model are the optional AI tools — AI Deep Review and Prompt Studio — and they run only when you click the button. There are no prompt-bearing background calls or automatic prompt submissions. The decision to send prompt text to the optional AI route is an explicit one you make.
Your browser
the prompt you built
You click “Run”
prompt text passes only after this action
CasePrompts server route
checks access and rate limits
Model provider
processes the prompt
Skip the AI tools and the path simply never opens. The free builder lives entirely in the first box.
03
Retention depends on provider and hosting accounts
When you run an optional AI feature, the prompt text is sent to Anthropic's API. CasePrompts does not intentionally persist the request or model response. Provider and hosting-infrastructure logs may still be created and retained under the configured services' account and operational settings. The configured API organization's terms control provider retention. Anthropic’s current commercial retention documentation describes standard deletion within 30 days with stated exceptions; its zero-data-retention documentation explains scope and model-specific limits. There is no document-upload feature; do not paste case documents, privileged facts, or confidential client information into the prompt fields.
Your prompt
sent on your click
CasePrompts server route
forwards the request
Model provider processes it
generates a response
Result back to you
through the server route
Application persistence: not intentional. Provider and infrastructure logs may still be retained under their configured account terms.
Where the models run
Not every feature needs a model, and not every model needs the cloud. We match each job to the lightest place it can run — and we always tell you which is which.
CITE structure checks
CITE Check is rule-based—it checks for the four CITE blocks, a source-location instruction, and a direction to report when support is not located. Prompt text stays on the device and no model-provider request is involved. If optional allowlisted product analytics are enabled, CasePrompts may record non-sensitive event metadata.
AI Deep Review & Prompt Studio
These optional tools use Claude through a cloud API. Prompt text is sent only when you choose to run one of the features. CasePrompts does not intentionally persist the request or result; provider and hosting-infrastructure logs may still be retained under their configured account terms.
Controls we can describe today.
These are repository-backed implementation boundaries—not a certification, compliance approval, penetration test, or claim that independent legal, privacy, or security review has occurred.
- Browser/server boundary
- Provider credentials are confined to server-side optional AI routes. The free builder and local CITE Check do not use them or make prompt-bearing model requests.
- Fail-closed private beta
- Cloud experiments require explicit configuration, invited access, a signed HTTP-only session cookie, and rate limits. AI routes refuse requests when required limiting is unavailable.
- Bounded product events
- Custom events use a finite, typed vocabulary; an unknown or sensitive property rejects the whole event. Site analytics load only when configured; production use is not approved until URL, referrer, IP, retention, and deletion handling are verified for the exact account and snippet.
- Release verification
- Automated checks cover common secret patterns, preview indexing, and runtime header contracts. Passing checks do not prove deployed behavior or security compliance; production still requires live verification.
Read the editorial and validation method or report a security issue.
What the current public tools do not do
- Upload or store case files—Ask your files reads them in your browser tab, and nothing is sent or kept
- Intentionally persist a copy of prompts you build or submit to optional AI features
- Send prompt-field contents to a CasePrompts API when you use the free builder
- Place prompt-field contents in setup or share URLs
Built so you don't have to choose between useful and careful.
The principle is simple: assemble the free prompt locally, make any prompt-bearing cloud call a deliberate choice, and disclose when provider and infrastructure terms apply. You still own the duty to verify every output against the record. These boundaries reduce unnecessary transmission of prompt text; they do not decide whether particular information is appropriate to enter or establish legal, confidentiality, or security compliance.