Prompt Library and Versioning

This standard governs the library of instructions: prompts: that direct the Company's AI services, and how a prompt version is created, tested, activated, and withdrawn. It is used by the AI Operations Lead, who approves activation; by AI Operators, who prepare and test new versions; and by the Clinical Reviewer, who approves any version affecting Pawzly's triage or safety wording. A prompt is behaviour: a one-line change can alter what thousands of customers are told. The library exists so that the exact words serving at any moment are known, tested before they serve, and reversible in seconds.

Roles and duties

• AI Operations Lead: approves activation and withdrawal of every prompt version.
• AI Operator: prepares candidate versions, runs evaluations, and records results.
• Clinical Reviewer: reviews the safety wording of every Pawzly prompt version and the hard cases described below, before activation.

The prompts in service

The library holds one active version per feature. The Pawzly features and the rules each active prompt carries are

1. Conversational reply. The assistant is Pawzly, the shopping and pet-care helper for All About Pawz in Memphis, giving brief, practical answers. It is not a veterinarian. It shall never diagnose. On any of the fixed emergency signs: trouble breathing, collapse, seizures, poisoning, bloat, repeated vomiting, blood, inability to urinate: it shall direct the customer to contact an emergency veterinarian immediately. It shall never invent, assume, or embellish pet names, breeds, weights, or details; it may refer only to pets explicitly on file, and where none are on file it names no specific pet. It shall never mention suppliers, distributors, or internal sourcing. It shall never claim it cannot do something the site does.
2. Product selection. The step answers in a fixed structured form only: whether the shopper would benefit from a product, a search query of 2 to 5 plain words a store search would match (including dog or cat where known), and a short reason. A malformed answer produces no products for that turn rather than an error shown to the customer. This step is the least costly call in the flow and the one most able to be wrong quietly: an inaccurate query returns irrelevant products, not a failure signal: which is why prompt changes to it still require evaluation.
3. Grounding rewrite. The preliminary reply is rewritten to under 110 words against the products actually returned: every safety point is kept; no product, brand, or product line may be invented; only products in the returned list may be named, and only where relevant; where none were returned, the product type is described and nothing is named; pet details shall not be invented; the no-diagnosis boundary is kept. This step is the second line of defence: the answering step is constrained a second time by the real product list.
4. Triage. The triage prompt directs urgency classification and returns the urgency level, red flags, whether the customer must be directed to a veterinarian, and the reply text. Because it carries the highest stakes on the platform, it is held to the same library, evaluation, and approval rules as every other prompt, and its current active version shall always be retrievable from the library in full.

Operators reading a prompt shall read its prohibitions first. Every prohibition in the active Pawzly prompts records a past failure or near-miss that was corrected with a sentence; removing a prohibition because its reason is not obvious is prohibited.

The classroom tutor's prompts are held in the same library under the same versioning rules, at the classroom's lower risk level.

Library rules

1. One record per version. The version number increments; a superseded version is never edited. The text of a version that served traffic is fixed in practice: a change is made by creating the next version, never by altering the serving one. Evaluation and incident records refer to a version, and they must always resolve to the exact words that served.
2. Each record states its feature: conversational reply, product selection, grounding rewrite, triage, or classroom tutor: which is the key joining the version to invocation records.
3. Status follows: prepared, then in evaluation, then active, then withdrawn. One active version per feature. Activation and withdrawal of the predecessor happen in the same action, so there is never doubt about what is serving.

Testing a prompt change

Both halves are required before activation

1. Evaluation. The candidate runs against the standard dataset under Evaluations and shall meet or beat the active version on every safety measure: no-diagnosis compliance, urgency accuracy on the reviewed triage sessions, escalation correctness, and grounding fidelity (every named product exists in the returned list). Better prose with any safety regression is refused, without exception.
2. Adversarial reading by a human. The Clinical Reviewer and an AI Operator put the hard cases to the candidate and read the answers: a bloated dog described casually; a male cat straining to urinate mentioned as an aside; a shopper asking whether a product is safe for a puppy where the product is age-restricted. The evaluation measures what it was built to measure; the human reading exists to catch what no one thought to measure.

Activation and rollback

A new version is prepared, moves to evaluation while testing runs, and becomes active only on the AI Operations Lead's approval after evaluation passes and human reading is complete. Rollback is a status change, not a deployment: withdraw the new version and reactivate the previous one. That speed is the safety argument for the library: a prompt regression in clinical classification is reversed before the next customer finishes typing. Customer data used in testing is limited to what the Privacy Policy permits, and test cases drawn from real sessions are used only in reviewed, corrected dataset form under Evaluations.

Controls

No prompt is edited outside the library process. The active version for each feature shall be identifiable at any moment from the library alone. Any discrepancy between the library's active version and the words actually serving is an incident under AI Incident Handling.

Records

This standard produces prompt version records, evaluation results for each candidate, adversarial-reading notes, and activation and withdrawal records with approver and time. Version records are retained permanently.

Escalation

A prompt change proposed for urgent safety reasons may be fast-tracked by the AI Operations Lead with the Clinical Reviewer, with evaluation completed retrospectively within one business day; the fast-track and its reason are recorded. Other questions go to support@aapawz.com; questions about the clinical boundary in prompt wording go to the Clinical Reviewer, and legal questions about disclosure wording to legal@aapawz.com.

Review cadence

The AI Operations Lead reviews the active versions and library completeness monthly, and the Clinical Reviewer re-reads the active Pawzly prompts quarterly.