
Universal AI Prompt Engineer
Turn rough prompts into professional, review-ready instructions for AI coding agents — without leaving VS Code.

Table of contents
Why this extension
Short prompts produce vague work. "Fix the save button" gives an agent no scope, no evidence
requirements and no definition of done, so you get a plausible-looking change that misses the
actual bug.
Prompt engineering fixes this, but rewriting every request by hand is slow and inconsistent.
This extension gives you a panel inside VS Code that takes the rough version and returns a
structured instruction — with the right emphasis for the task at hand (implementation, debugging,
review, design, gameplay, autonomy) — in a single keystroke.
- No prompt library to memorise. Nine modes encode the structure for you.
- Bring your own model. Any OpenAI-compatible
/chat/completions endpoint works.
- No secrets in your settings. The API key lives in VS Code
SecretStorage.
- Works where you already work. Selection, clipboard, panel, sidebar, status bar, palette.
Features
|
|
| Nine enhancement modes |
Quick, Professional, Developer, Architect, Debugging, Code Review, UI/UX, Game Development, Autonomous Agent |
| Dedicated panel |
Two-column editor/result layout that follows your VS Code theme (dark, light, high contrast) |
| Selection aware |
Ctrl+Alt+P enhances the selected text; the sidebar pre-fills it into the panel |
| Clipboard aware |
Enhance whatever you copied from any chat or document |
| Configurable output |
Show in panel, copy to clipboard, or replace the selection in place |
| Secret-safe |
API key stored in SecretStorage, never in settings.json, never echoed back to the UI |
| Honest errors |
Provider failures are surfaced verbatim (HTTP status + provider message), never replaced with a fake success |
| Guard rails |
HTTPS-only endpoints, no credentials in URLs, 20,000 character cap, 90 second timeout |
| Zero runtime dependencies |
Only vscode and Node's built-in fetch |
Screenshots

The Prompt Engineer panel
| Dark theme |
Light theme |
 |
 |
Quick start
You need your own API key. The extension does not bundle a key, host a model, or pay for
provider usage. Any OpenAI-compatible provider works.
1. Install
- VS Code Marketplace:
Ctrl+Shift+X → search Universal AI Prompt Engineer → Install
- From a VSIX file:
code --install-extension universal-ai-prompt-engineer-0.2.0.vsix
2. Connect your provider (once)
Press Ctrl+Alt+O or click $(sparkle) Prompt Engineer in the status bar.
Expand Provider settings and fill in:
| Field |
Example |
| HTTPS API endpoint |
https://api.openai.com/v1 |
| Model |
gpt-4o-mini |
| API key |
your key |
Click Save provider settings. The header pill changes to Provider connected.
The key is written to VS Code SecretStorage. Leaving the key field blank on later saves keeps
the stored key unchanged.
3. Enhance
- Type or paste your rough prompt.
- Pick a mode.
- Press Enhance prompt (or
Ctrl+Enter).
- Press Copy and paste the result into Copilot Chat, Claude, Cursor, or any agent.

The nine modes
| Mode |
What it optimises for |
Good for |
| Quick Enhance |
Tight, specific wording — no extra structure |
one-line asks, follow-up messages |
| Professional Enhance |
Context, scope, deliverables, success criteria |
general work, documentation, mixed tasks |
| Developer Mode |
Conventions, error handling, security, focused tests |
implementing a feature in an existing codebase |
| Expert Architect |
Requirements, boundaries, data flow, trade-offs, staged delivery |
system design, migrations, cross-cutting change |
| Debugging Mode |
Reproduction, evidence, root cause, minimal fix, regression check |
"it sometimes breaks" reports |
| Code Review Mode |
Prioritised findings by severity — no unsolicited rewrites |
reviewing a PR or a diff |
| UI/UX Mode |
User goals, layout, responsive behaviour, accessibility, states |
screens, flows, design-system work |
| Game Development Mode |
Mechanics, controls, feedback, assets, performance, playtesting |
gameplay and engine work |
| Autonomous Agent Mode |
Inspect → plan → implement → verify → report actual results |
long-running agent tasks |
Every mode preserves your intent. The system prompt explicitly forbids inventing technologies,
requirements or results, and tells the downstream assistant to inspect the project when context
is unknown.
Ways to run an enhancement
| Entry point |
How |
Source text |
| Panel |
Ctrl+Alt+O → paste → Enhance |
what you typed |
| Selection |
select text → right-click → Enhance Selected Prompt or Ctrl+Alt+P |
the selection |
| Clipboard |
Ctrl+Shift+P → Enhance Prompt from Clipboard |
the clipboard |
| Sidebar |
Activity Bar → Prompt Engineer → click a mode |
current selection (pre-filled) |
| Palette |
Ctrl+Shift+P → AI Prompt Engineer: … |
input box |
| Status bar |
click $(sparkle) Prompt Engineer |
opens the panel |
Commands
| Command |
Default binding |
AI Prompt Engineer: Enhance Selected Prompt |
Ctrl+Alt+P (with selection) |
AI Prompt Engineer: Enhance Prompt from Clipboard |
— |
AI Prompt Engineer: Open Prompt Engineer Panel |
Ctrl+Alt+O |
AI Prompt Engineer: Set API Key |
— |
AI Prompt Engineer: Clear API Key |
— |
AI Prompt Engineer: Quick Enhance |
— |
AI Prompt Engineer: Professional Enhance |
— |
AI Prompt Engineer: Developer Mode |
— |
AI Prompt Engineer: Expert Architect Mode |
— |
AI Prompt Engineer: Debugging Mode |
— |
AI Prompt Engineer: Code Review Mode |
— |
AI Prompt Engineer: UI/UX Mode |
— |
AI Prompt Engineer: Game Development Mode |
— |
AI Prompt Engineer: Autonomous Agent Mode |
— |
Internal commands used by the sidebar and webview: openPromptPanelWithMode,
enhanceSelectionWithMode, enhanceClipboardWithMode.
Keyboard shortcuts
| Shortcut |
Action |
Ctrl+Alt+O |
open / reveal the panel |
Ctrl+Alt+P |
enhance the current selection |
Ctrl+Enter (in panel) |
run the enhancement |
All bindings can be changed in Keyboard Shortcuts (Ctrl+K Ctrl+S) by searching for
universalAIPromptEngineer.
Settings
| Setting |
Default |
Description |
universalAIPromptEngineer.apiEndpoint |
https://api.openai.com/v1 |
HTTPS base URL of an OpenAI-compatible API. Must use HTTPS, no credentials, query or fragment. |
universalAIPromptEngineer.model |
gpt-4o-mini |
Model name accepted by that provider. |
universalAIPromptEngineer.defaultMode |
professional |
Mode used by selection, clipboard and palette commands. |
universalAIPromptEngineer.outputMode |
panel |
panel, copy, or replace. |
universalAIPromptEngineer.showStatusBar |
true |
Show or hide the status bar item. |
The API key is not a setting. Use the panel, Set API Key, or Clear API Key.
Output modes
outputMode |
Behaviour |
Best for |
panel |
opens/reveals the panel and shows the result |
reviewing and iterating |
copy |
writes the result to the clipboard and shows a notification |
fast paste into a chat |
replace |
overwrites the current editor selection with the result |
rewriting prompts inside files |
How it works

Request lifecycle
sequenceDiagram
autonumber
actor U as You
participant W as Webview / Command
participant E as extension.ts
participant S as SecretStorage
participant C as Workspace config
participant P as aiProvider.ts
participant A as AI provider (HTTPS)
U->>W: rough prompt + mode
W->>E: postMessage { type: "enhance", text, mode }
E->>S: read API key
E->>C: read endpoint + model
E->>P: requestPromptEnhancement()
P->>P: getCompletionUrl() (HTTPS validation)
P->>A: POST /chat/completions (Bearer key)
A-->>P: choices[0].message.content
P-->>E: enhanced prompt
E->>W: postMessage { type: "result" }
U->>W: Copy
W->>E: postMessage { type: "copy" }
E-->>U: clipboard
Design decisions
- One choke point. Every entry point funnels into
enhancePromptFromText(), so validation,
errors and output routing behave identically no matter how the user started.
- Text never touches settings. Only the endpoint and model are persisted in configuration;
the key is read at request time from
SecretStorage.
- Strict URL handling.
getCompletionUrl() refuses non-HTTPS URLs, embedded credentials,
query strings and fragments, then appends /chat/completions if it is missing.
- No silent fallbacks. A failed request throws with the provider's message; the UI shows it.
There is no template-based "enhancement" pretending to be a model response.
- CSP + nonce. The webview ships a strict Content-Security-Policy and a per-load nonce.
- Theme native. The panel uses
--vscode-* CSS variables, so dark, light and high-contrast
themes all work without a second stylesheet.
Privacy and security
| Data |
Where it goes |
| API key |
VS Code SecretStorage on your machine; sent only in the request header to your endpoint |
| Endpoint + model |
your user settings.json |
| Prompt text |
sent to the HTTPS endpoint you configure, and nowhere else |
| Enhanced result |
kept in memory/clipboard/panel; not logged or uploaded |
- The extension never reads Copilot or other extensions' chat input.
- No telemetry, analytics, or background requests.
Clear API Key deletes the stored secret immediately.
- Provider pricing, data retention and privacy are between you and your provider.
Report a vulnerability privately as described in CONTRIBUTING.md.
Requirements
- VS Code 1.90 or newer.
- An OpenAI-compatible provider that accepts
Authorization: Bearer <key> and the
/chat/completions request shape (OpenAI, OpenRouter, Groq, Together, a local gateway, …).
- Network access to that provider over HTTPS.
Install from source
git clone https://github.com/mzainulabideendev/Universal-AI-Prompt-Engineer.git
cd Universal-AI-Prompt-Engineer
npm install
npm run compile
Then press F5 (a launch.json with the extensionHost configuration is included) to open the
Extension Development Host window.
npm run watch # incremental type-check while editing
npm run package # builds a .vsix
Testing
Unit tests
npm test
Compiles with tsc and runs the Node test runner over tests/:
| Suite |
Covers |
tests/aiProvider.test.cjs |
URL building, HTTPS/credential rejection, request shape, mode instructions, provider errors, empty completions, input limits |
tests/promptEngine.test.cjs |
mode catalogue integrity, labels, descriptions, fallback behaviour |
These tests inject a fake fetch, so they need no API key and no network.
Manual smoke test
Do this in the Extension Development Host window:
| # |
Step |
Expected |
| 1 |
Enable without a saved key |
Error banner + Set API Key / Open Settings actions |
| 2 |
Save a valid key |
Pill shows Provider connected |
| 3 |
Enhance in Professional |
Structured prompt with scope and success criteria |
| 4 |
Same text in Debugging |
Repro → evidence → root cause → verified fix |
| 5 |
Select text → Ctrl+Alt+P |
Result appears in the panel |
| 6 |
Copy text → Enhance Prompt from Clipboard |
Result appears |
| 7 |
Click a sidebar mode |
Panel opens with that mode preselected |
| 8 |
outputMode = copy |
Result goes to the clipboard |
| 9 |
outputMode = replace with a selection |
Selection is replaced |
| 10 |
Endpoint http://… |
"must use HTTPS", no request is sent |
| 11 |
Wrong model name |
AI provider request failed (HTTP 4xx): … |
| 12 |
Clear API Key |
Pill returns to API key required |
| 13 |
Reload the window |
Endpoint/model persist, key still reported as saved |
Regenerating the README images
The panel screenshots are rendered from the real webview markup in src/extension.ts, not from a
hand-drawn mock, so they cannot drift from the shipped UI:
node docs/preview/build-images.cjs # all images
node docs/preview/build-images.cjs --only panel
node docs/preview/build-images.cjs --only diagrams
It needs Chrome or Edge installed (headless rendering only).
Packaging and publishing
npx vsce package # writes universal-ai-prompt-engineer-<version>.vsix
npx vsce login mzainulabideen # Marketplace personal access token
npx vsce publish # or: npx vsce publish minor
Check the shipped file list first with npx vsce ls. The VSIX contains only
dist/, resources/, docs/images/, README.md, LICENSE and package.json.
Bump version for every publish — the Marketplace does not allow overwriting a version.
Project structure
.
├── src/
│ ├── extension.ts activation, commands, tree views, panel webview, output routing
│ ├── aiProvider.ts URL validation, chat-completions request, response parsing
│ └── promptEngine.ts mode catalogue (id, label, description, icon, tooltip)
├── tests/ node:test suites compiled from dist/
├── dist/ compiled JavaScript (what VS Code loads)
├── resources/ Marketplace icon (png) and Activity Bar icon (svg)
├── docs/
│ ├── images/ README banners, diagrams and screenshots
│ └── preview/ webview preview generator + sidebar mock
├── visualstudio/ optional Visual Studio 2022 port (separate VSIX)
└── .github/workflows/ CI: compile + tests on every push and PR
Troubleshooting
| Symptom |
Cause / fix |
| No API key is configured |
Save a key in Provider settings, or run Set API Key |
| must use HTTPS… |
Endpoint must start with https:// and contain no ?query, #fragment or user:pass@ |
| AI provider request failed (HTTP 401) |
Key is invalid, revoked, or belongs to another provider |
| HTTP 404 / model not found |
Model name does not exist at that endpoint |
| timed out after 90 seconds |
Slow model or network — pick a faster model or retry |
| The prompt is too long |
Prompts are capped at 20,000 characters |
| Panel is blank |
Developer: Open Webview Developer Tools → check the console for CSP errors |
| Commands missing |
Developer: Show Running Extensions → confirm the extension is activated |
Edits to src/ have no effect |
Run npm run compile (F5 loads dist/, not src/) |
| Light theme looks wrong |
Update to 0.2.0+, which uses --vscode-* theme variables instead of hard-coded colours |
Roadmap
- [ ] Optional streaming responses
- [ ] Prompt history and saved snippets
- [ ] Per-mode system prompt overrides
- [ ]
@prompt-engineer chat participant
- [ ] Open VSX release for VSCodium and Gitpod
Ideas are welcome as issues.
Contributing
Bug reports, mode improvements and prompt-design feedback are all useful.
See CONTRIBUTING.md for the workflow, code style and checklist.
npm install
npm run compile
npm test
License
MIT © mzainulabideen
| |