PromptCraft docs
Everything you need to get from download to optimized prompts in under 5 minutes. No account required. No configuration wizards.
Installation
PromptCraft ships as a signed and notarized .dmg. Apple's Gatekeeper will verify it on first launch — no approval hoops required.
Go to the Download page and grab the latest release. The file is named PromptCraft-1.0.0.dmg.
Open the .dmg and drag PromptCraft.app into your /Applications folder. Eject the disk image when done.
Double-click to launch. macOS will confirm you want to open an app from an identified developer — click Open. PromptCraft will guide you to System Settings → Privacy & Security → Accessibility to grant the single read-only permission it needs.
The PromptCraft icon appears in the right side of your menu bar. Click it to open the app window, or use the global keyboard shortcut directly from any app.
PromptCraft is signed with Developer ID and notarized by Apple. It passes Gatekeeper on first launch without any bypass steps. If macOS blocks it, go to System Settings → Privacy & Security and click Open Anyway.
Keyboard shortcut
The shortcut is the core interaction model. You never need to click the menu bar icon — just press the keys from anywhere on macOS and the window appears over your current work.
Default global shortcut — toggle PromptCraft from anywhere
To change the shortcut: open PromptCraft, click the gear icon in the header bar, navigate to Shortcut. Click the field and press your new key combination. PromptCraft will register the new shortcut system-wide immediately.
If the shortcut conflicts with another app, PromptCraft will warn you. Common conflicts: ⌘⇧Space is sometimes used by Raycast or Alfred. Pick something like ⌃⌥Space or ⌘⌥P if needed.
The 7-stage pipeline
Every prompt you type passes through a seven-stage processing pipeline before it reaches any AI model. The whole sequence runs locally on your Mac in under 200ms.
Optimization styles
Styles are the top-level control over how the Prompt Assembler shapes your request. Switching style does not change your input — it changes the assembly template the pipeline uses, producing a completely different, equally precise output.
Balanced structure that works across most task types. Good for everyday requests where you do not need domain-specific framing.
Adds role specification ("You are a senior…"), explicit constraint enumeration, format requirements (JSON / TypeScript / annotated code), and unit test directives. Ideal for code, architecture, and technical review tasks.
Academic framing, source awareness instructions, structured section requirements (background / methodology / findings). Optimized for literature synthesis and evidence-based analysis.
Narrative structure, audience tone calibration, SEO awareness, headline variants. Built for blog posts, landing copy, social, and long-form editorial.
Analytical frameworks (SWOT, first-principles, root cause), hypothesis testing structure, structured conclusions with supporting evidence. For business analysis, data interpretation, and decision support.
Formal register, citation structure requests, methodology sections, abstract framing. For academic writing, research proposals, and peer-review style output.
Open-ended framing, voice latitude, inspiration triggers, sensory language instructions. For fiction, poetry, screenwriting, and imaginative brainstorming.
Define your own template using the template editor. Custom styles appear in the pill row alongside built-ins. Build a house style once and reuse it across every request.
Verbosity modes
Verbosity controls how much the Post Processor asks the model to elaborate. It is applied per-request and does not persist between sessions unless you set a default in Settings.
Strips preamble, requests direct answers, limits response length. Use when you know exactly what you need and want it immediately — quick refactors, short explanations, single-purpose answers.
Preserves full context, standard response length. The pipeline uses this unless you select otherwise. Works for 90% of everyday use cases.
Requests step-by-step reasoning, examples for edge cases, and full elaboration of any assumptions. Use for learning, debugging complex problems, or generating comprehensive documentation.
Context Engine
The Context Engine is a fully local semantic memory system. Every optimization is embedded on-device using Apple's NaturalLanguage framework and stored in a private SQLite database. Nothing ever leaves your Mac. Over time, the engine builds a map of your projects, vocabulary, and working patterns, and automatically retrieves relevant context for each new prompt.
How it works
When you optimize a prompt, PromptCraft generates a 512-dimensional embedding of your input and output text using a local NLP model — no API call, no internet. That embedding is stored alongside structured metadata: persons, projects, environments, technical terms, and temporal markers extracted by the Entity Extractor stage.
For each new optimization, the engine computes cosine similarity against all stored entries and retrieves the top matches above the relevance threshold (default 0.65). The matched context is injected into the Prompt Assembler stage without you writing any background.
Project clusters
PromptCraft automatically groups your past prompts into named project clusters. Cluster names are derived from the most frequent entity in each group (project name, environment, or technical term) — not random IDs. You can rename any cluster from the context panel inside the app.
Reclustering happens automatically in the background every time 10 new entries are indexed. Clusters with fewer than 5 entries are hidden from the display until they grow.
| Setting | Default | Description |
|---|---|---|
| Context Engine enabled | On | Toggle memory collection on or off |
| Relevance threshold | 0.65 | Minimum cosine similarity to include an entry |
| Max entries per prompt | 1,000 | Total stored context entries before oldest are pruned |
All embeddings and context data are stored in a private SQLite database at ~/Library/Application Support/PromptCraft/context.sqlite. The database never syncs to iCloud, never leaves your machine, and is excluded from backups by default.
Templates
Templates let you define reusable prompt structures with named placeholders. When you apply a template, PromptCraft fills in the placeholders and passes the assembled text through the full 7-stage optimization pipeline before showing the result.
Creating a template
Click the grid icon in the app header, then click New Template.
Use {{placeholder}} syntax for variable parts. Example: Explain {{concept}} to a {{audience}} in {{tone}} tone.
Select the template from the popover. PromptCraft shows a form for each placeholder. Fill them in and press Optimize.
Exporting and importing templates
Templates can be exported as a .promptcraft-templates JSON file and imported on another machine. Use Settings → Templates → Export All. The file includes all template text, categories, and metadata, but no personal context data.
Inline overlay
The inline overlay lets you optimize text directly inside any macOS application — Xcode, VS Code, Notion, Notes, Mail, Terminal, or any text field. Select the text you want to optimize, wait for the overlay to appear, and press Enter. The original text is replaced with the optimized version.
Setup
The inline overlay requires the Accessibility permission already granted during installation. Enable it in Settings → Behavior → Inline Overlay.
| Setting | Default | Description |
|---|---|---|
| Inline overlay enabled | On | Show overlay when text is selected |
| Activation delay | 500ms | How long to wait after selection before showing overlay |
| Excluded apps | None | Comma-separated app bundle IDs to skip |
If you use an app where text selection triggers unexpected behavior, add its bundle ID to the excluded list. To find a bundle ID, run osascript -e 'id of app "AppName"' in Terminal.
Watch folder
The watch folder monitors a directory on your Mac. When a new .txt file appears in that folder (up to 100 KB), PromptCraft optimizes its contents automatically and copies the result to your clipboard. A system notification confirms when optimization completes.
Setup
Go to Settings → Automation → Watch Folder and toggle it on.
Click Choose Folder and pick any accessible directory. A common pattern is a desktop folder named to-optimize.
Optionally assign a specific optimization style to the watch folder. Files dropped here will always use that style, regardless of your current selection in the app.
PromptCraft tracks processed filenames to avoid reprocessing the same file twice. To reprocess, delete the file and drop a fresh copy.
Local API
PromptCraft exposes a local REST API on port 9847. Any tool, script, or workflow can send a prompt and receive back the optimized result. The Alfred workflow (included in the download) uses this API. You can build custom integrations with Raycast, shell scripts, or any HTTP client.
Enable the API
Go to Settings → Behavior → Local API Server and toggle it on. Copy your bearer token from the same panel.
Endpoint
Authorization: Bearer YOUR_TOKEN
Content-Type: application/json
{
"text": "your raw prompt here",
"style": "engineering",
"verbosity": "concise"
}
Alfred workflow
The Alfred workflow is included in the PromptCraft download folder. Import it into Alfred (Powerpack required), set PROMPTCRAFT_TOKEN and optionally PROMPTCRAFT_PORT in the workflow variables, then type pc your prompt text in Alfred to copy an optimized result to clipboard.
| Setting | Default | Description |
|---|---|---|
| Local API enabled | Off | Start the local HTTP server |
| Port | 9847 | Listen port (configurable) |
| Rate limit | 10 req/min | Per-token request cap |
Export formats
After optimization, the output row shows a copy button and a download menu. PromptCraft supports five clipboard formats and four file export destinations.
Clipboard formats
| Format | Description |
|---|---|
| Plain Text | Raw optimized text, no formatting |
| Markdown | Wrapped in a fenced code block with a prompt language tag |
| Claude (XML) | Wrapped in <prompt> XML tags, compatible with Claude Projects and API calls |
| ChatGPT | Formatted as a ChatGPT-compatible user message |
| GitHub Issue | Formatted with a title line and body, ready to paste into a new issue |
File export destinations
| Destination | Output file |
|---|---|
| Cursor Rules | .cursorrules — paste into your project root for Cursor IDE |
| Claude Project Instructions | system-prompt-claude.md — upload to a Claude Project |
| ChatGPT Custom Instructions | chatgpt-instructions.txt — paste into GPT settings |
| Raw System Prompt | system-prompt.txt — plain text for any provider |
Explain mode and diff view
Enable Explain mode in settings to see a brief note after each optimization explaining what the pipeline changed and why. Enable Diff view to see a word-level diff highlighting every insertion and deletion between your input and the optimized output.
License activation
After purchase, a license key in the format PC-XXXX-XXXX-XXXX-XXXX is delivered to your email within 60 seconds via our automated system. Check your spam folder if it does not arrive.
The key is 19 characters including hyphens, in the format PC-XXXX-XXXX-XXXX-XXXX. Copy it to your clipboard.
Click the gear icon in the PromptCraft header bar. Navigate to License and click Enter License Key.
Paste your key and click Activate. PromptCraft contacts our license server to verify and register the machine. Activation takes under 2 seconds with an internet connection.
To transfer to a new machine: go to Settings → License → Deactivate on the old Mac first, then activate on the new one. If you lost access to the old Mac, email hello@promptcraft.app with your order ID and we will reset the activation count manually within 24 hours.
Ollama setup
Ollama lets you run open-weight models locally — no API key, no internet, no cost per token. PromptCraft detects a running Ollama instance automatically.
Download from ollama.ai and install. Ollama runs as a background service on localhost:11434.
Open Terminal and run:
Other good options: mistral, qwen2.5-coder, deepseek-r1. The first pull downloads the weights — subsequent launches are instant.
Open Settings → AI Provider → Local. Set the base URL to http://localhost:11434 and select your downloaded model from the dropdown. PromptCraft will list all locally available models automatically.
With Ollama selected, your prompts never leave your Mac. The entire pipeline runs on-device. No network traffic, no token costs, no rate limits. Works on flights, submarines, and cafes with bad Wi-Fi.
Cloud routing
Cloud subscribers route prompts to frontier models without managing API keys. The proxy handles authentication, rate limiting, and retries automatically.
To enable: go to Settings → AI Provider → Cloud. Select your preferred provider and model. PromptCraft sends your assembled prompt to our Cloudflare Workers proxy at promptcraft-cloud-proxy.ozzydev.workers.dev over HTTPS. The proxy authenticates with the provider and streams the response back.
We log only: a hash of your license key (not the key itself), the provider name, HTTP status code, and response latency. No prompt content is ever logged or stored. The hash is used solely for rate-limit enforcement.
FAQ
Yes. PromptCraft works independently of any chat interface. It optimizes your prompt in a separate window, then you paste the result wherever you want — Claude.ai, ChatGPT, a terminal, or a custom app.
The Context Engine reads your active application name and selected text to inject relevant context automatically. The permission is read-only — PromptCraft cannot type, click, or control your Mac. You can revoke it at any time in System Settings.
Yes. Go to Settings → AI Provider → API Key and paste your key. PromptCraft connects directly to the provider without routing through any proxy.
Yes. PromptCraft launches with a 7-day free trial, no credit card required. The full pipeline runs in trial mode — the only limitation is the trial banner in the window.
Nothing. Pro licenses are perpetual and cover all 1.x updates. The license key you activated today will keep working after every update.
Pro licenses activate on 1 machine. Cloud licenses activate on up to 2. You can deactivate from one Mac and activate on another at any time.
Changelog
- Initial public release of PromptCraft
- 7-stage pipeline with full local execution
- 8 optimization styles including custom template support
- 3 verbosity modes (Concise, Balanced, Detailed)
- Ollama integration with automatic model discovery
- PromptCraft Cloud proxy with 5 AI providers
- Pro and Cloud license tiers via Stripe
- Signed and notarized for macOS 14 Sonoma and later