Copilot Optimizer AIGet better Copilot answers automatically. Fix prompts. Improve answers. Learn over time. Installation Launch VS Code Quick Open ( Ctrl+P), paste the following command, and press enter.Copied to clipboard |
Copilot Optimizer AI
A smart layer between you and GitHub Copilot that improves results automatically — no manual tuning required. 🆕 Native chat:
|
| Setting | Default | What it does |
|---|---|---|
copilotOptimizer.tokenBudget |
400 |
Accuracy ↔ token-savings preset (Max savings · Balanced · Accuracy-first · Maximum accuracy). Lower = more compression, higher = more context. |
copilotOptimizer.contextFiles |
3 |
How many top-ranked files to attach as real code. 0 = names only. |
copilotOptimizer.contextCharBudget |
6000 |
Max characters of file content sent to the model. |
copilotOptimizer.model |
"" |
Preferred Copilot model (e.g. Claude Opus 4.8, GPT-5.5). Empty = use the model picked in the chat dropdown. |
copilotOptimizer.enabled |
true |
Master on/off for optimization. |
Quick Start
| Action | How |
|---|---|
| Optimize a prompt | Ctrl+Shift+O — type prompt → press Enter → paste into Copilot |
| Quick Send (killer feature) | Select text → Ctrl+Alt+Enter → paste into Copilot |
| Right-click | Select any text → right-click → Quick Send |
That's it. The engine handles strategy, context, and compression for you.
Example
Before:
please help me fix the null pointer exception in UserService
After (Ctrl+Alt+Enter):
fix: null pointer UserService
mode: debug
focus: root cause
minimal response
Copilot answers faster, more precisely, with less noise.
Why it works
Bad prompts → vague answers → retries → wasted time.
The optimizer:
- Detects intent — is this a fix, a feature, a question?
- Compresses — removes noise, keeps signal
- Adds context — ranks your codebase and injects relevant nodes
- Learns — adapts strategy based on what worked before
Features
| Feature | What it does |
|---|---|
| Intent detection | Recognizes fix, feature, refactor, explain, test, docs |
| Adaptive compression | Removes filler words while preserving meaning |
| Context graph | Ranks your workspace files by relevance to the prompt |
| Strategy memory | Remembers what strategy worked best per intent |
| Fuzzy cache | Reuses cached strategies for similar contexts |
| Confidence scoring | Detects low-confidence results and retries with wider context |
| Output shaping | Tells Copilot how to format the answer |
| Learning feedback | Gets smarter every time you use it |
Commands
| Command | Shortcut | Description |
|---|---|---|
| Improve Prompt | Ctrl+Shift+O |
Type a prompt → optimized result is copied |
| Quick Send | Ctrl+Alt+Enter |
Selected text → optimize → copy instantly |
| Show Dashboard | Command Palette | Stats, learned strategies, context graph |
| Score Response | Command Palette | Rate a Copilot response to improve learning |
| Rebuild Graph | Command Palette | Re-scan workspace context |
How the learning works
Every time you click "✓ Great answer" or "↩ Regenerated" after using the extension, the system:
- Updates which strategy works best for that intent
- Adjusts context node weights (which files matter for which prompts)
- Caches the winning strategy for similar future prompts
After ~10 uses, it becomes noticeably smarter for your specific codebase.
Privacy
All optimization, scoring, and learning run locally on your machine. No prompts, no code, and no codebase context are sent anywhere except your configured Copilot provider.
- Zero telemetry
- Zero external API calls
- All learning data stored in VS Code's per-user global storage on your machine
Release Notes
1.1.0
- Native
@optimizerchat participant — write naturally in Copilot Chat; it rewrites your prompt, sends it to the model, streams the answer, and auto-scores the response. - Real file-content context — injects the actual code of the most relevant workspace files (not just file names) for far more accurate answers.
- Slash commands —
/debug/fix/refactor/explain/generate/raw/dashboard. - Settings — token-budget preset, context file/character budgets, preferred model, and a master on/off switch.
- Working on/off toggle — the status-bar switch now actually enables/disables optimization.
- Fix — learning data now persists in writable per-user global storage (previously failed silently when installed from the Marketplace).
1.0.0
- Adaptive learning engine (v6)
- Fuzzy cache matching (Jaccard similarity)
- Node weight decay (30-day half-life)
- Confidence scoring with automatic context expansion
- Quick Send (
Ctrl+Alt+Enter) — zero-friction optimization - RELEASE_MODE: stable 2-variant search
Better prompts → better answers → fewer retries.