Glossary of terms

Terms used across the Prompt Assay app, docs, and help center. Alphabetical within groups. If a term isn't here and you think it should be, email support.

Core concepts

Prompt

The atomic unit of work. A prompt has a title, description, content, prompt type, optional target model, and intent metadata. Always belongs to exactly one workspace.

Version

An immutable snapshot of a prompt's content at a point in time. Every save creates a new version tagged with what caused the change (manual edit, AI improvement, restore, branch, etc.).

Fragment

A reusable snippet of prompt text with optional variable slots. Referenced from prompts via ${fragment:id}. Good for shared preambles, policy blocks, or output-format directives.

Folder

An organizational label for prompts within a workspace. A prompt can be in at most one folder.

Tag

A color-coded label for prompts. A prompt can have multiple tags. Tags belong to the workspace they were created in.

Test suite

A set of test cases attached to a prompt, with expected criteria and a pass threshold. Used for regression testing and model comparison.

Rubric

A list of weighted criteria the judge scores an output against. Each criterion has a description (what to check) and a numeric weight (how much it matters).

Multi-tenancy

Workspace

A container for prompts, folders, tags, fragments, BYOK keys, billing, and membership. Two kinds: personal (one per user, single-member) and team (created manually, up to 15 members).

Personal workspace

Auto-created on signup. Exactly one member forever (you). Cannot be invited into, cannot be deleted except via account deletion, cannot be converted to a team.

Team workspace

Manually created. Supports owner / admin / member roles. Up to 15 members on Team tier. Owner can transfer ownership, delete, or leave.

Active workspace

Whichever workspace is currently displayed in your sidebar and used for all your queries. Tracked in a signed cookie bound to your user, so another person logging in on the same browser can't inherit your active workspace.

Role (owner/admin/member)

Your role inside a specific workspace. Strict total ordering: owner > admin > member. Enforced in multiple layers so no single bug can unintentionally open a permission.

LLM + BYOK

BYOK (Bring Your Own Key)

A per-workspace API key for Anthropic, OpenAI, or Google. Required at every tier. Stored encrypted at rest; spend flows to your provider account on every call.

Target model

The LLM a prompt is *designed for* at runtime. Metadata on the prompt. Drives the token budget bar and linter warnings. Independent of which model actually runs the prompt in the playground. You can critique a target-GPT-4.1 prompt using Claude as the workbench model.

Workbench model

Your personal default model for the six AI pair panels (Brainstorm, Critique, Improve, Rewrite, Compare, Convert). The Judge that powers eval suites and regression sweeps also honors this preference. Stored as a user preference. Distinct from a prompt's target_model. Convert filters the picker further to models that meet its hardware floor (≥64k output and streaming).

Judge

An AI call that scores a test case's output against a rubric across six dimensions (clarity, completeness, structure, technique usage, robustness, efficiency) on a 0-10 scale, producing a weighted overall score. Judge is a separate eval-suite mechanism · NOT one of the six AI pair panels. Honors your Workbench Model at every tier. Free / Solo grandfathered workspaces see an informational cost warning for expensive models but execution is never blocked.

AI Chains

A sequence of AI Pair actions (e.g., Critique → Improve, or Critique → Improve → Rewrite) that run back-to-back from a single button click. Each step's output is passed as context to the next. A cost preview modal shows estimated tokens and USD before the chain fires. Enable via Settings → Profile → AI Chains. Chain configuration is stored on your device only. Convert is origin-only and not part of any chain.

Convert

The sixth AI pair panel. One-shot decomposition of a plain prompt into a complete Agent Skill bundle (SKILL.md plus optional scripts/ and references/), with a preview gate before anything lands in your library. Available on both the Prompts and Skills workbenches. BYOK-only (not on the demo allowlist), rate-limited to 10 runs per minute, blocked during admin impersonation, and restricted to models with ≥64k output tokens and streaming.

Dimension pinning

A per-user preference (Settings → Profile → Critique dimensions) to skip specific Critique dimensions that don't apply to your workflow, e.g., robustness on a one-shot internal tool. Pinned dimensions are excluded from the radar chart, dimension list, and chain context. Stored on your device; no data is lost when un-pinning.

Improve vs Rewrite

Two related but distinct refactoring modes. Improve returns up to 8 atomic before/after edits that you cherry-pick one-by-one (additive, reversible). Rewrite returns a full refactored version of the whole prompt that you accept or reject wholesale via a diff view (wholesale, but reviewable). Rewrite is NOT a from-scratch rewrite. It reads your current content + any critique context and produces a coherent rebuild. Use Improve for surgical fixes; use Rewrite when multiple weaknesses need to be fixed in a coordinated way.

Primary / valid flags on a BYOK key

Each stored key is marked either primary (the one actually used) or non-primary (during rotation). If a provider returns an auth error, the key is automatically marked invalid and further calls fail fast until you rotate, preventing accidental lock-out from a stale key.

Billing

Tier

A workspace's billing plan: free, solo, team, or enterprise. Controls monthly LLM call cap, max members, API access, and SAML SSO availability. Per-workspace, not per-user.

Over-quota

Flag set when a workspace crosses its tier's monthly LLM call cap. Blocks every LLM call until the next billing period resets the counter.

Free default

Newly created workspaces have no explicit subscription until someone completes a Stripe checkout for them. Until then, the free-tier caps apply.

Billing period

The monthly window your subscription renews on. Stripe sets it on your first checkout; all usage counts and caps reset on the next renewal date. For workspaces without a subscription, the window is a rolling 30 days from now, not a calendar month.

Legal + compliance

GDPR data export

An account-wide ZIP of everything you own (prompts, versions, fragments, settings, usage logs). Requested from Account settings → Danger zone. Limited to 1 per 24 hours.

Privacy policy acceptance

When we publish a privacy-policy update flagged as a major update, every authenticated session is redirected to an acceptance page until you explicitly accept the new version. Minor edits (typos, clarifications) don't trigger the gate.

DPA (Data Processing Agreement)

Contract between Prompt Assay and an Enterprise customer specifying the terms under which personal data is processed. Published in the Legal section.