Copilot Token Optimizer - Context & Prompt Advisor

Analyse your GitHub Copilot prompts and open editor tabs before sending — get actionable suggestions to reduce token usage and improve prompt quality.

Core analysis is 100% local. Optionally enhance suggested prompts with a local Qwen AI model via Ollama.

How to Use

1. Click the 🧠 Copilot Advisor icon in the Activity Bar
2. Paste your prompt → click 🔍 Analyse (or Cmd+Shift+J / Ctrl+Shift+J)
3. Context tab — see which tabs to close/open to save tokens
4. Prompt Tips tab — expand any tip for details and one-click fixes
5. Click ▶ Open in Copilot Chat to send the optimized prompt directly

What It Checks

Context Analysis — Are your open tabs costing you?

Every file tab you have open gets sent to Copilot as context — even files you're not working on. This eats into your token budget silently.

• Large files eating tokens — Flags files over 500 lines that inflate your context window
• Irrelevant open tabs — Detects tabs unrelated to your current task (e.g. a CSS file open while you're fixing a backend bug)
• Expensive file types — XML, YAML, lock files, and logs are verbose and waste tokens fast
• Test files you don't need — If you're not writing tests, open test files just add noise
• MCP tools running in background — Extensions like Atlassian, GitKraken, and Pylance register tools that silently consume ~250 tokens each per request
• Wrong Copilot mode — Recommends Ask, Agent, or Edit mode based on what you're actually trying to do

Prompt Quality — Is your prompt wasting tokens?

A poorly written prompt can cost 2-5x more tokens and give worse results. The extension checks 12+ common mistakes:

• Filler phrases — "Can you please help me..." wastes tokens. Just say what you need
• Multiple tasks in one prompt — "Fix the login AND refactor the database AND add tests" should be 3 separate prompts
• No file reference — Without #file:YourFile.php, Copilot guesses which file you mean and loads extra context
• No output format — Without "return only the code, no explanation", Copilot writes paragraphs of explanation you don't need
• Vague language — Words like "something", "stuff", "things" force Copilot to guess what you want
• Ambiguous pronouns — "Fix it" or "change this" — Copilot doesn't know what "it" refers to
• No stopping condition — Without "only modify this file", agents may keep going — editing extra files, adding commits you didn't ask for
• Missing success criteria — Tell Copilot what "done" looks like: "Done when the timeout is 30s and tests pass"
• Mixing research + implementation — "Find the bug AND fix it AND write tests" should be separate conversations
• Agent tasks without safety rails — Autonomous agent prompts need confidence gates and budget caps to prevent runaway token usage

Local AI Setup (Optional)

Enhance suggested prompts with Qwen 2.5 Coder running locally via Ollama:

# Install Ollama
brew install ollama

# Pull the model
ollama pull qwen2.5-coder:7b

# Start Ollama (runs on http://127.0.0.1:11434)
ollama serve

The extension auto-detects Ollama. When available, suggested prompts show 🤖 Qwen + rules. Without Ollama, you get 📋 Rule-based suggestions — still fully functional.

Settings

Keyboard Shortcuts