How to use Prompt Lab

Quick start

Quality scoring

The editor automatically scores your prompt on five dimensions:

• Role — detected if you include "you are," "act as," or similar persona cues
• Task — action verbs like "write," "generate," "explain"
• Format — output format specified (JSON, markdown, list, table, etc.)
• Constraints — guardrails like "must," "avoid," "limit"
• Context — sufficient background (80+ characters)

Score color: green (4-5/5), yellow (2-3/5), red (0-1/5).

Golden Responses

When you get an output you're happy with, click Pin Golden. This saves it as a benchmark. Future iterations will show a side-by-side comparison against your golden reference, so you can tell if quality is drifting.

Layout modes

Toggle between three editor layouts using the tabs below the header:

• Editor — full-width prompt editing
• Library — browse your saved prompts
• Split — editor on the left, library on the right

Browsing & filtering

• Search — filters by title and tags in real time
• Tags — click a tag chip to filter (Writing, Code, Research, Analysis, Creative, System, Role-play, Other)
• Collections — custom folders you create; click a collection name to filter
• Sort — Newest, Oldest, Most Used, or Manual (drag to reorder)

Per-entry actions

• Load — sends the prompt back into the editor
• Copy — copies the enhanced version to clipboard
• Expand — shows full content, version history, test cases, golden response, notes, and metadata
• Share — generates a URL with the prompt encoded in it (no server needed)
• Add to Composer — inserts as a block in the Compose tab
• Rename / Edit / Delete — manage entries

Version history

Every time you save changes to an existing prompt, the previous version is preserved. Expand any entry to see its timeline:

• Browse prior versions with timestamps and change notes
• Toggle Show Diff for inline word-level comparison
• Click Restore to revert to any previous version

Import & Export

At the bottom of the library panel:

• Export — downloads your entire library as a .json file
• Import — loads prompts from a JSON file. Supports both Prompt Lab's native format and external schemas (fields like prompt, description, category are auto-mapped)

How to run a test

History & recent runs

The A/B Test tab keeps two collapsible sections:

• Recent Runs — individual test executions with provider, model, and latency
• History — full experiments with variant labels, dates, and the winning variant

Three-panel layout

• Library panel (left) — your saved prompts. Click + or drag to add a block to the canvas
• Canvas (center) — your working composition. Drag blocks to reorder. Click X to remove
• Preview (right) — live preview of the final composed prompt with markdown headers and section separators

On mobile, these collapse into Library | Canvas | Preview toggle tabs.

Output actions

• Copy All — copies the entire composed prompt
• Send to Editor — loads the composition into the Editor tab for enhancement
• Clear — removes all blocks from the canvas

Managing pads

• New — creates a new named pad (prompted for a name)
• Rename — change the active pad's name
• Delete — remove the active pad (disabled if only one remains)
• Switch between pads using the sidebar list on the left

Toolbar

• Date — inserts a formatted date separator (e.g., "── Fri, Mar 14, 2025 ──")
• Download — exports the pad as a .txt file
• Copy — copies all content
• Clear — empties the pad (with confirmation)

Auto-save

Content saves automatically as you type. The status indicator in the bottom-right shows:

• Saving... — debounced write in progress
• Saved — last save succeeded (with timestamp)
• Save failed — storage error (rare; usually means localStorage is full)

User-defined variables

Wrap any placeholder in double braces: {{role}}, {{language}}, {{context}}. When you enhance, Prompt Lab detects all {{ }} patterns and shows a form to fill in values before sending.

Ghost variables (auto-resolve)

These built-in variables fill themselves at runtime — no input needed:

Providers

Select your active provider from the dropdown. Each provider has its own fields:

• Anthropic — API key (sk-ant-...), model (default: claude-3-5-sonnet-20241022)
• OpenAI — API key, model (default: gpt-4o)
• Google Gemini — API key, model (default: gemini-2.0-flash)
• OpenRouter — API key, model (default: anthropic/claude-3.5-sonnet)
• Ollama — base URL (default: localhost:11434), model. Click Refresh Models to auto-detect loaded models

Testing your connection

Click Test Connection to validate your API key and model. A success or error message confirms the result. Keys are stored in localStorage and never leave your browser.

PII detection

If Prompt Lab detects sensitive content in your prompt (API keys, passwords, tokens), it shows a warning with three options:

• Redact & Send — replaces sensitive data with placeholders
• Send Anyway — proceeds without redaction
• Cancel — aborts the operation

Global

Scratchpad

Scratchpad shortcuts are limited to local text actions. Browser-reserved shortcuts like ⌘T, ⌘W, ⌘[, and ⌘] are intentionally not overridden.

Command palette

Press ⌘K to open the palette. Type to filter, then select:

• Enhance Prompt, Save Prompt, Clear Editor
• Go to Editor / Composer / A/B Test / Pad
• Toggle Light / Dark mode
• Export Library, Open Settings
• Show Keyboard Shortcuts