Set up Agent Context System with GitHub Copilot as a skill

Install Agent Context System as a GitHub Copilot skill to get persistent project memory, automatic updates, and cross-session context in every Copilot session.

GitHub Copilot is the recommended way to install the Agent Context System. Installing through the official Copilot skill registry means you receive updates automatically whenever the system improves — no manual file copying required. After setup, Copilot reads your AGENTS.md through a .github/copilot-instructions.md directive that the init script creates for you.

Install via skill registry

1. Add the skill

   Run the following command in your project root to install the Agent Context System from the Copilot skill registry:

   npx skills add AndreaGriffiths11/agent-context-system
   

   Installing through npx skills add pins you to the official registry entry, so you get updates to the system automatically. This is the only install method that provides automatic updates.

2. Run the init script

   After the skill installs, run the init script to wire up your project:

   bash .agents/skills/agent-context-system/scripts/init-agent-context.sh
   

   The script creates .agents.local.md from the template, ensures it is gitignored, and adds a directive to .github/copilot-instructions.md telling Copilot to read AGENTS.md at session start.

3. Edit AGENTS.md

   Open AGENTS.md and fill in your project specifics: name, stack, build commands, and the patterns, boundaries, and gotchas that are true of your codebase. The placeholders in the template are starting points — replace them with real knowledge from your project.

   This is the highest-leverage edit you will make. The more accurate your AGENTS.md, the more useful every Copilot session becomes.

Install manually (no package manager)

If you prefer not to use npx skills, you can copy the Copilot skill file directly into your repository.

1. Copy the skill file

   Copy github-copilot/SKILL.md from the Agent Context System repository to your project:

   mkdir -p .github/skills/agent-context-system
   cp /path/to/agent-context-system/github-copilot/SKILL.md \
      .github/skills/agent-context-system/SKILL.md
   

   If you have not cloned the repository yet, clone it first to a temporary location:

   git clone https://github.com/AndreaGriffiths11/agent-context-system.git /tmp/acs
   mkdir -p .github/skills/agent-context-system
   cp /tmp/acs/github-copilot/SKILL.md .github/skills/agent-context-system/SKILL.md
   rm -rf /tmp/acs
   
2. Add the copilot-instructions.md directive

   Create or update .github/copilot-instructions.md with a directive pointing Copilot to AGENTS.md:

   Read AGENTS.md before starting
   
3. Edit AGENTS.md

   Copy AGENTS.md to your project root and fill it in with your project specifics. Keep it under 120 lines — that constraint is intentional.

Manual installation does not receive automatic updates. If the Agent Context System releases improvements to the Copilot skill, you will need to re-copy the file yourself.

How Copilot reads AGENTS.md

After init, your .github/copilot-instructions.md contains a directive that tells Copilot to read AGENTS.md before starting any task. Copilot loads copilot-instructions.md at the start of every Chat session, so your project knowledge is always in context — no manual retrieval required.

Session logging

Copilot Chat does not expose session-end hooks, so the Agent Context System cannot log sessions automatically. To save what happened in a session to .agents.local.md, tell Copilot at the end of your conversation:

log this session

Copilot will propose a session log entry covering what changed, what worked, what did not work, and any patterns worth remembering. Review the proposal and approve it before Copilot writes to .agents.local.md.

The session log is where patterns begin their journey into AGENTS.md. After enough sessions, Copilot will flag recurring patterns in the scratchpad’s “Ready to Promote” section. You decide when to move them. See the knowledge flow guide for how promotion works.