Master the AIChemy workflow.
Learn the execution-oriented improve pipeline, provider capability model, web and extension contract, and the exact controls that turn rough intent into production-ready prompts.
Start with the real job, constraints, audience, and risk so the improved prompt preserves purpose instead of only sounding cleaner.
A living map of the improve engine
AIChemy starts by reading the job behind the prompt so the result preserves what you actually need done.
Hover or tap any signal to inspect how the guide translates into real product behavior.
Overview
What AIChemy does now
AIChemy improves prompts for downstream AI systems with an execution-oriented pipeline. Instead of only polishing wording, it classifies request intent, applies provider-aware rendering, and returns a normalized result shape that both the web app and extension understand.
Intent-aware improvement
AIChemy classifies whether your input is a task, a prompt draft, or a meta-prompt about improving another prompt. Explicit controls override inference when you need precision.
Execution-first scoring
The active improve flow prioritizes execution readiness over prompt polish with scores for intent preservation, actionability, tool readiness, and verification readiness.
Shared result contract
Web and extension consume the same normalized result shape: improvedPrompt, techniques, explanation, warnings, metadata, execution score, and legacy score compatibility.
Quick start
Get productive quickly
- 1
Sign in with Google and open the dashboard workspace.
- 2
Connect at least one provider in Settings. On the web app, provider keys stay client-side. In the extension, provider settings are stored locally and encrypted.
- 3
Paste the prompt you want to improve, pick a model, choose an agent, and optionally set segment or prompt effort.
- 4
Use advanced controls when you want to force intent: input mode, downstream target, provider preference, desired output, target audience, project context, or risk level.
- 5
Run the improvement, inspect the improved prompt plus execution score, then save it to history if you want to reuse it later.
Mastery map
Think like a power user
Prompt mastery relay
Move through the core operating loop and watch the active phase update like a command center instead of reading a static checklist.
Frame the job
Write the real task, not a polished slogan. AIChemy works best when it can see the desired outcome, audience, constraints, and risk.
Frame the job
Write the real task, not a polished slogan. AIChemy works best when it can see the desired outcome, audience, constraints, and risk.
Choose the right controls
Use agent, model, segment, and prompt effort for broad steering. Add advanced controls only when the output needs non-negotiable intent.
Inspect the result
Read warnings, execution score, techniques, and explanation. Treat them as a map of what the system preserved, inferred, or downgraded.
Reuse what works
Save strong outputs to history, start from templates for repeated work, and refine with project context instead of rebuilding from scratch.
Workspace
Controls that matter
Always-visible controls
Advanced improve controls
Providers
Current provider model
OpenAI
Strong general, coding, and reasoning coverage. Current improve flow uses the Responses API and supports long-output recovery when the provider reports max-output incompletion.
Anthropic
Reasoning and agentic workflows with provider-specific thinking support. Improve rendering uses Anthropic-style structured sections instead of an OpenAI-shaped prompt.
Multimodal and OAuth-capable provider surface. Improve requests can use Google OAuth where available and use provider-specific rendering instead of a universal prompt template.
Mistral / Groq / Meta
Supported through the same normalized capability matrix. Behavior parity is defined at the AIchemy layer, with explicit downgrade warnings when a requested capability is not available.
/api/models. Specific model names and availability can change over time, so treat the settings panel and public model API as the source of truth.Extension
How the extension fits in
Popup flow
- 1Select text or use the popup entry path to create a pending prompt.
- 2If you are authenticated, the popup can improve immediately through /api/improve/stream.
- 3Streaming responses expose progress and final normalized results; pre-stream failures return structured JSON errors.
- 4Improved prompts and related metadata can be saved into extension history and synced into the web dashboard view.
Storage and safety
The extension stores sensitive local state through Web Crypto AES-GCM helpers. It validates sender identity for internal messages, restricts external messaging to AIChemy domains, and uses server-signed extension sessions instead of unsigned local tokens.
History & templates
Saved work and reuse
History
Web and extension history preserve the normalized improve result shape, including warnings, execution score, input mode, downstream target, and provider metadata.
Templates
Templates help you start faster, then advanced controls add stronger constraints when the task needs richer intent and output requirements.
Security
Important behavior and limits
- Your web-side AI provider keys stay client-side.
- The extension stores sensitive settings locally with encryption helpers rather than plain local storage values.
- Improve requests still use a marker-based response contract internally so the server can parse provider output consistently.
- When a model cannot honor the requested thinking or tool posture, AIChemy returns downgrade warnings instead of pretending full parity exists.
Troubleshooting
Common issues
No providers available
Open Settings and connect a provider first. The workspace and extension both rely on a configured model before an improve request can run.
The model is available in one place but not another
The public model list comes from /api/models. If an admin changes model availability, make sure the public catalog and admin source stay aligned.
The improved prompt is too generic
Set explicit advanced controls: choose the input mode, set a more specific downstream target, and add desired output / project context so the renderer has real constraints to preserve.
A request times out or downgrades capability
AIChemy emits structured warnings when the selected model cannot fully honor the requested reasoning or tool posture. Switch models or reduce requested effort if needed.