Dyph is a prompt engineering workspace where you assemble system prompts from modular, versioned blocks — then run, compare, and snapshot the results. No subscriptions, no paywalls.
Every detail was designed around one goal: iterate on prompts faster and smarter.
Each prompt block holds up to 9 saved text variants for the same semantic slot. Switch between them instantly with a single tap — no duplicate cards, no copy-pasting. This is your built-in A/B testing layer: same block position, different formulations, one click apart.
Break your system prompt into named, reorderable blocks. Each block is a self-contained semantic unit — enable or disable it without touching the rest. Only enabled blocks reach the model.
Save a named snapshot of your full workspace state: all prompt blocks, all parameter values. Restore any snapshot instantly — perfect for safe experimentation without losing what works.
Every run is logged with timestamp, full prompt, and model response. Click "Details" to see the exact request that was sent — parameters, full system + user prompt, token counts.
Before you run, hit Preview to see the exact combined text that will be sent as the system prompt, followed by your user prompt. No guessing what gets assembled from which blocks.
The controls adapt to the selected model: Reasoning Effort & Verbosity for GPT-5 / o-series; Temperature, Top-P, Frequency & Presence penalty for GPT-4o. Plus Seed and JSON Mode for any model.
Your API keys are securely stored and never exposed. Only you have access to your credentials — Dyph keeps them safe so you can focus on building prompts.
No installs, no config files. Just open the app and start building.
Go to Settings → Credentials, paste your OpenAI key, and save. It stays encrypted in your browser — we never touch it.
Hit "New Project", give it a name and optional description. Each project is its own workspace with independent prompts, outputs, and snapshots.
Add blocks for each part of your instructions — role, tone, output format, constraints. Fill in up to 9 variations per block. Drag to reorder, toggle to disable. Check Preview to see the final text.
Type your User Prompt, pick a model and parameters, hit Run. The response lands in Output. Tweak a block or parameter, run again. When something works — Save Snapshot.
Switch models without leaving your project. Parameters adapt automatically.
More providers — Anthropic, Google, xAI — coming soon