Prompt Optimizer for VS Code

A local prompt optimizer, semantic cache, and conversational AI agent � all running inside VS Code without leaving your editor.

?? Fully local � zero telemetry. All prompt processing, token counting, semantic caching, and secret scanning run entirely on your machine. No prompts, code, file contents, or metrics are ever sent to any external server. The only network request is the final Copilot call you explicitly trigger.

🚀 Interactive Visual Onboarding Guide

Markdown renderers (GitHub, Marketplace, VS Code preview) strip <script> and most <style> blocks for security, so an animated HTML walkthrough cannot run inside README.md itself. The extension ships the full interactive guide as a webview instead:

Command Palette → Prompt Optimizer: Open Guide

(or run prompt-proxy.openOnboarding from the keyboard shortcut editor)

What the live guide gives you that this static page cannot:

• A 3-frame animated simulator — workspace indexing → prompt typed in Copilot Chat → optimized prompt + cache hit, with neon node-graph filtering down to just the files your prompt actually touched.
• An interactive knowledge graph of 10 sample files with edges that light up and dim based on the active context.
• A memory pipeline with three conveyor belts (Ingestion → Storage → Retrieval) and a lane-flow diagram showing how the optimized prompt reaches Copilot.
• Keyboard shortcuts: Space play/pause, R reset.
• Full Catppuccin Mocha palette with neon glow, cyberpunk grid, animated cursor, ripple effects, and typing carets — all rendered as real DOM, not source code.

The static file lives at vscode-extension/media/onboarding.html — you can also open it directly in any browser if you want to share it.

What you'll see (static preview)

┌─────────────────────────────────────────────────────────────────┐
│ Prompt Optimizer · Onboarding                          [P spin] │
├─────────────────────────────────────────────────────────────────┤
│ [How it works] [Memory & storage] [Privacy & commands]          │
├──────────────────┬──────────────────────────────────────────────┤
│ ① Index          │   ●─────●        app.ts                      │
│   workspace      │  /│     │\                                   │
│ ② Type a prompt  │ ● │     │ ●     auth.ts ✦ (lit up)           │
│ ③ Optimized +    │  \│     │/                                   │
│   cache hit      │   ●─────●        token.ts ✦ (lit up)         │
│                  │   others dimmed                              │
│ [ ▶ Play ]       │   ⚡ +71.5× Token Savings · cache hit         │
└──────────────────┴──────────────────────────────────────────────┘

What it does

Prompt Optimizer sits between you and Copilot. Before your prompt reaches the model it:

1. Compresses filler words, softeners, and redundant phrasing
2. Checks the local semantic cache � reuses a prior result if a similar prompt was already answered
3. Packs workspace context � active file, open editors, diagnostics, recent saves
4. Estimates token cost before the request is sent
5. Calls Copilot and streams the answer directly in the sidebar (Agent mode) or hands you the optimized prompt to review first (Optimize mode)
6. Detects SDLC workflow intent from slash commands, built-in rules, custom skills, and file context
7. Remembers the conversation across turns per workspace so follow-up pronouns ("fix it", "add tests for that") resolve correctly
8. Lets you enable and edit bundled SDLC agents directly from the Agents button in the sidebar

Modes

Select the mode from the Mode dropdown in the sidebar or via the status bar item ($(robot) Proxy [Agent]).

Switch mode any time by:

• Clicking the status bar label ($(robot) Proxy [Agent]) ? QuickPick
• Changing the Mode dropdown in the sidebar
• Typing @promptoptimizer /mode agent (or optimize / direct) in Chat

SDLC modes and agent skills

Prompt Optimizer 2.1.0 includes a built-in SDLC mode layer that can frame the optimized prompt with a role and checklist before it is sent.

Built-in modes

Slash-triggered modes available out of the box:

• /plan
• /arch
• /code
• /test
• /review
• /security
• /qa
• /devops
• /docs
• /pr
• /full
• /bug-fix
• /refactor

If you do not type a slash command, the picker can still activate a mode by scoring prompt intent, keywords, requires tokens, filePatterns, and priority from the registered skills.

Custom skills

Workspace skills live in .promptoptimizer/skills/*.md. User-global skills can also be loaded from PROMPT_OPTIMIZER_SKILLS_DIR.

Supported frontmatter fields include:

• id, label, readOnly
• slashAliases, intentPatterns, keywords
• requires, filePatterns, priority, tags
• rolePreface as a scalar or block (|)
• checklist as frontmatter items or markdown bullets under ## Checklist

Skill files hot-reload on the next prompt. The parser tolerates comments, block scalars, quoted regex strings, and inline or block arrays. If a skill fails to load, diagnostics are available from the command palette.

Agents button in the sidebar

The Agents quick action opens a multi-select picker for the bundled SDLC agent library.

• Tick an agent to enable it for the current workspace.
• Untick an agent to disable it.
• Click the pencil icon to enable-and-open the workspace copy for editing.
• Click reset to restore the bundled definition over local edits.

Enabled agents are copied into .promptoptimizer/skills/ so you can customize them like any other workspace skill.

Sidebar panel � Prompt Optimizer Control

Open it from the Chat sidebar or press the status bar item.

+----------------------------------+
�  Prompt Optimizer       Local cache  �
�  Mode  [ Agent ? ]               �
� +------------------------------+ �
� � Type your prompt here�      ?� �  ? send button (icon, like Copilot)
� +------------------------------+ �
�  ? Alerts (secrets, errors)      �
�  [ Open @promptoptimizer ]           �
�  [ Use optimized ] [Copy] [Docs] �
�                                  �
�  +-- Analysis table ----------+  �
�  � Cost � $x  � Saved � n tok �  �
�  � Tok  � n?m � Est.  � n out �  �
�  � Cache� semantic 82%        �  �
�  � Price� in $x + out $y      �  �
�  +-----------------------------+ �
�  Refinements � Optimized prompt  �
�  � use nouns � [compressed text] �
�                                  �
�  +-- Copilot response --------+  �
�  � streaming�                  �  �
�  +-----------------------------+ �
+----------------------------------+

Send button

The circular ? button inside the textarea behaves like the Copilot send button � hover shows the current mode action ("Run Agent � optimize + call Copilot").

Chat participant � @promptoptimizer

Type in the VS Code Chat panel:

@promptoptimizer refactor the auth middleware to use async/await

Every message is automatically optimized. Copilot's answer streams back in chat.

Commands

Conversation memory

Up to 12 turns per workspace are remembered. Back-references resolve automatically:

Turn 1: "refactor the auth middleware to async/await"
Turn 2: "now add unit tests for it"
         ? proxy injects "[Continuing from: 'refactor�']" before sending

Analysis table (Optimize mode)

After analyzing a prompt, the result card shows a compact table:

Below the table, Refinements and the Optimized prompt sit side by side.

Secret detection

If your prompt contains what looks like an API key, token, or private key header, a ?? alert appears in the panel before anything is sent. Patterns detected:

• OpenAI keys (sk-�)
• Anthropic keys (sk-ant-�)
• GitHub tokens (ghp_�, ghs_�)
• AWS access keys (AKIA�)
• PEM private key headers
• Generic password=, token=, api_key= assignments

Semantic cache

The local SQLite cache stores every prompt you send and builds embeddings for semantic similarity. On future prompts it checks for:

• Exact match � returns the cached optimized version instantly
• Semantic match (= 68% cosine similarity) � returns and boosts confidence score
• Miss � optimizes fresh, writes to cache

The cache is seeded on activation from your git log, Copilot chat history, README, package.json, and AI instruction files (.github/copilot-instructions.md, AGENTS.md, etc.).

Manage via command palette:

• Prompt Optimizer: Show Cache Statistics
• Prompt Optimizer: Clear Semantic Cache

Privacy & Security

To verify: open the extension source (out/extension.js) or the engine source (engine/dist/). Search for http, fetch, axios, request � you will find zero outbound calls outside of the Copilot API.

Command palette

Configuration