Copilot Prompt Analyzer

You have already written the spec. It is buried in your chat history.

Every prompt you have sent to GitHub Copilot Chat is sitting on your disk — the requirements, the conventions, and every correction you had to make. This extension reads all of it, groups it by topic, and distils each area into one reusable prompt you can drop straight into .github/prompts/.

Runs on your machine, through the Copilot subscription you already have. No API key, no account, no telemetry.

Watch the narrated demo

The one that matters: working prompts

Open the Areas view, pick an area — auth, database, UI — and press Build working prompt. It reads every request you ever made in that area and produces a single document.

The default style is Portable: it generalises every specific instance into the rule behind it and names no repository, service or file path, so the result applies to any project including one you have not started yet.

• Role — what the assistant is being asked to do whenever it works here
• Rules — the standing requirements, numbered and checkable
• Conventions — how you want work done
• Do not — mined from the corrections you had to issue
• Definition of done — the checks to run before claiming completion
• Ask me first — decisions you have historically wanted to make yourself

The Do not section is usually the surprise. Those rules came from every time you typed "no, not like that" and then watched the reason evaporate when the session closed.

Switch to Project-anchored when you want the opposite: the stack, services and file layout of the work these requests came from, as a spec for continuing it.

Save it and it becomes something you use tomorrow:

Everything else

Grouped by topic. Prompts are classified into areas — UI, API, data, auth, security, infra, testing, performance, debugging and more. Don't like the taxonomy? Describe a different one in plain English and it regroups.

Ask your own history. "What did I keep getting wrong about auth?" "Extract every bug I reported, with dates." Answers are grounded in whichever prompts your filters select, and cite them by id.

Seven reports in the Insights tab. Three run entirely on your machine and cost nothing:

Improve this prompt. Open any prompt and get a rewrite plus an explanation of what changed.

Getting started

1. Click the Copilot Prompt Analyzer icon in the activity bar. Your history is read immediately — no setup, no export.
2. Press Classify with AI and approve model access. Results are cached by prompt text, so later runs only classify what is new.
3. Explore Areas, Prompts, Timeline and Insights. Filter by area, workspace, model, mode, date or free text.
4. Press Expand for the full-width dashboard in an editor tab. Both views share one data source and stay in sync.

Model, reasoning and context

Nothing here is a hardcoded list.

• Model — enumerated from your Copilot plan, each shown with the context window the provider reports.
• Reasoning — VS Code publishes no reasoning metadata: LanguageModelChat exposes only id, name, vendor, family, version and maxInputTokens, and modelOptions is an untyped passthrough. So Detect asks the provider directly, trying candidate option keys and effort values with one-token requests and keeping only what is accepted. The dropdown is populated from that result and stays empty when a provider genuinely refuses. Cached per model.
• Context — budget options are computed from the model's own window, so a 32k model and a 1M model both get sensible choices. The characters-per-token ratio is measured with countTokens() rather than assumed.

Open any working prompt or AI report and the footer shows the real token cost of what will be sent, before it is sent.

Commands

Settings

Privacy

Stays on your machine. Chat history is read from local files and never uploaded. Classifications and generated documents are stored locally. Three of the seven reports never make a network call at all. No telemetry, no analytics, no account.

Worth knowing. Classification and synthesis send prompt text to a language model through your Copilot subscription, under its terms. Your prompts contain code, paths and anything you pasted into chat — filter to a project first if some of it is sensitive. Prompt text is truncated to copilotPromptAnalyzer.maxPromptChars before it is sent. Exports are as sensitive as the original chats.

Requirements

VS Code 1.95 or later, and an active GitHub Copilot subscription with chat.

Only chat history still on disk is visible — VS Code prunes old sessions, so running the extension periodically preserves more of it going forward.

Not affiliated with GitHub or Microsoft.