Getting Started

Set up once. ContextManager works in the background from that point forward.

Step 1 — Create a Project

  1. Click the 📖 book icon in the Activity Bar
  2. Click + to create a new project
  3. Name your project (e.g., “my-app”)
  4. Select the workspace folders that belong to this project
  5. Open the Dashboard (click the status bar item or run ContextManager: Open Dashboard)

Your project is now active. ContextManager will capture knowledge from all future Copilot interactions in this workspace.

ContextManager auto-discovers .github/copilot-instructions.md and README.md and includes them in project context.

Step 2 — (One-time) Register the Hook

To capture responses from all chat participants, register the VS Code agent hook once:

  1. Open the Dashboard → Settings tab
  2. Expand the 🪝 Agent Hooks section
  3. Click Install Hooks — this copies capture.ps1 to ~/.contextmanager/scripts/ and writes hooks.json to .github/hooks/

Without the hook, intelligence still accumulates from auto-capture during your active chat sessions. With the hook, every Copilot Chat, @workspace, and background agent response is also captured automatically.

Step 3 — Chat Normally

Just use Copilot as you normally would:

What does the AuthController class do?
How is the error handling pipeline structured?
Explain the retry logic in NetworkClient.

ContextManager captures intelligence silently in the background:

What’s captured How
Conventions LLM extracts coding patterns from responses
Tool Hints Learned search terms that work for your codebase
Working Notes Code relationships and insights discovered during exploration
Card Candidates High-confidence knowledge staged in the Card Queue for review

No commands. No setup. It just runs.

If you have multiple ContextManager projects, open the Dashboard → Sessions tab and bind each tracked chat session to the right project. Unbound sessions queue their captures until you assign them.

Step 4 — Review the Queue (Once a Week)

AI responses accumulate in the Card Queue. When you have a few:

  1. Open the Dashboard → Knowledge tab → 📬 Card Queue subtab
  2. Click 🤖 Distill into Cards — one LLM call synthesizes all queued items into card proposals
  3. Review proposals: click ✓ Add for ones worth keeping (or Approve All)
  4. If a similar card already exists, a merge picker appears — merge is the default action

You don’t need to review the queue every session. Weekly or per-feature review is plenty.

Step 5 — Sessions Start Informed

From now on, every new Copilot session automatically receives context two ways:

  • copilot-instructions.md managed block: ContextManager auto-syncs a managed section into your .github/copilot-instructions.md containing #ctx tool usage instructions and pinned card titles. VS Code always includes this file, so every agent starts informed.
  • #ctx tool: Available to all agents for on-demand search, list, learn, getCard, and explicit card queue review across all project knowledge.

The AI knows your architecture, conventions, and recent discoveries without you repeating yourself.

The Flywheel 

graph TD
    A[Work with Copilot normally] --> B[Knowledge captured in background]
    B --> C[Queue: review and approve cards]
    C --> D[Cards injected into every future session]
    D --> E[Copilot starts informed]
    E --> A

    style A fill:#1f6feb,stroke:#388bfd,color:#fff
    style B fill:#1158c7,stroke:#388bfd,color:#fff
    style C fill:#238636,stroke:#3fb950,color:#fff
    style D fill:#1158c7,stroke:#388bfd,color:#fff
    style E fill:#1f6feb,stroke:#388bfd,color:#fff

Each session builds on everything that came before it.

LM Tools Reference

5 Language Model Tools available to all agents:

#ctx                   — unified search, list, learn, getCard, queue review, distillQueue, clearQueue
#getCard               — read a specific knowledge card by ID
#saveCard              — save a new knowledge card
#editCard              — edit an existing card
#organizeCards         — organize cards into folders

If multiple ContextManager projects exist, add project:"Exact Project Name" or the exact project ID/root path to every LM tool call.

Optional: Custom Workflows

Once you have knowledge accumulating, you can automate recurring AI tasks:

  1. Open the Dashboard → Intelligence tab → Custom Workflows
  2. Click + New Workflow and define a prompt template using `` placeholders
  3. Choose a trigger — manual, auto-queue, or event-based (convention-learned, card-created, etc.)
  4. Set the output action — create card, update card, or append to a collector card

Workflows run against your project data and can either generate markdown with AI or save rendered templates directly. See Custom Workflows for details.

Next Steps

Features →