VS Code's GitHub Copilot — and the GitHub Copilot CLI — both look for custom agents, instructions, skills, and prompt files under %USERPROFILE%\.copilot\{agents,instructions,skills,prompts}. By default these stay local to a single machine and never sync.

CopilotAtelier solves that by storing the canonical files in a single, repo-derived folder (preferring OneDrive for cross-machine sync) and then linking the well-known ~/.copilot/* discovery folders to that target with NTFS junctions. Write an agent once, use it from both the VS Code Copilot chat extension and the Copilot CLI, on every machine.

No chat.*FilesLocations settings are written for agents, instructions, or skills — those three are auto-discovered by both clients via the junctions. Prompts are the exception: the VS Code chat extension does not auto-discover ~/.copilot/prompts, so the script writes a single chat.promptFilesLocations entry for that one path. The CLI auto-discovers it on its own.

Only one canonical location is populated per machine (no duplicate mirror). If a previous version of the script left a stale local mirror behind, the new run cleans it up automatically when OneDrive is present.

~/OneDrive/CopilotAtelier/       # Used when OneDrive is detected (preferred)
~/CopilotAtelier/                # Fallback — used only when OneDrive is not installed
├── Agents/          # Custom agents (.agent.md files)
├── Instructions/    # Custom instructions (.instructions.md files)
├── Skills/          # Agent skills (folders with SKILL.md)
├── Prompts/         # Prompt files / slash commands (.prompt.md files)
├── Keybindings/     # keybindings.json merged into your user profile
├── Setup-CopilotSettings.ps1   # Setup script
└── README.md        # This file

The folder name is derived from the repository clone, so renaming the clone renames the synced layout automatically.

The setup script configures the following in settings.json:

Discovery is junction-based for agents, instructions, and skills — the script does not write chat.agentFilesLocations, chat.instructionsFilesLocations, or chat.agentSkillsLocations because both the VS Code Copilot chat extension and the GitHub Copilot CLI auto-discover the well-known ~/.copilot/{agents,instructions,skills} paths. Prompts are the exception: VS Code Copilot Chat reads prompts only from %APPDATA%\Code\User\prompts and from paths listed in chat.promptFilesLocations (only the CLI auto-discovers ~/.copilot/prompts), so the script writes a single chat.promptFilesLocations entry for ${userHome}/.copilot/prompts via a merge that preserves any user-added prompt locations. The script copies the four customization folders to a single canonical target and creates NTFS junctions under %USERPROFILE%\.copilot\ so both clients see the same files:

%USERPROFILE%\.copilot\agents       --> <target>\Agents
%USERPROFILE%\.copilot\instructions --> <target>\Instructions
%USERPROFILE%\.copilot\skills       --> <target>\Skills
%USERPROFILE%\.copilot\prompts      --> <target>\Prompts

Where <target> is %USERPROFILE%\OneDrive\CopilotAtelier when OneDrive is installed, otherwise %USERPROFILE%\CopilotAtelier.

If one of the ~/.copilot\<name> folders already exists as a real directory:

• Empty → removed silently and replaced with the junction.
• Non-empty → the script prompts before deleting. On y, its contents are merged into the target (existing target files are not overwritten) and then the directory is removed and replaced with the junction. On n, the junction is skipped and a warning is printed.

Existing junctions are recreated on every run so they always point at the current target.

Note on model availability: Opus 4.7 requires Copilot Pro+, Business, or Enterprise. On other plans VS Code falls back to its default. To pin a different model, edit Setup-CopilotSettings.ps1 or override these two settings in your settings.json after running the script.

The setup script merges the bindings in Keybindings/keybindings.json into %APPDATA%\Code\User\keybindings.json. The merge is idempotent (match key: key + command + when), preserves user-added bindings, and creates a timestamped backup on every run.

1. Clone this repository (or sign into OneDrive if you already keep a synced clone there).
2. Open PowerShell and run the setup script from the cloned location, for example:

# From a local clone
& "<path-to-clone>\Setup-CopilotSettings.ps1"

# Or from a OneDrive-synced clone
& "$env:USERPROFILE\OneDrive\CopilotAtelier\Setup-CopilotSettings.ps1"

The script copies the contents of the clone into ~/OneDrive/CopilotAtelier/ when OneDrive is detected, or into ~/CopilotAtelier/ otherwise, then creates NTFS junctions at ~/.copilot/{agents,instructions,skills,prompts} pointing to that target so both the VS Code Copilot chat extension and the GitHub Copilot CLI pick up the same files. It also patches your VS Code settings.json and keybindings.json idempotently with timestamped backups. If a stale ~/CopilotAtelier/ mirror exists from a previous dual-copy run, the script removes it when OneDrive is now used. If a pre-existing non-empty folder is found at any of the ~/.copilot/* link paths, the script asks before deleting it (and copies its contents into the target first).

3. Restart VS Code.

• Agents: In Copilot Chat, check the agents dropdown — your custom agents should appear
• Instructions: Type /instructions in chat to see the Configure Instructions menu
• Skills: Type / in chat to see skills as slash commands
• Prompts: Type / in chat to see prompt files as slash commands
• Chat Customizations editor: Click the gear icon in the Chat view (or run Chat: Open Chat Customizations from the Command Palette) to see all registered agents, instructions, skills, and prompts in one place
• Debug logs: If customizations aren't being applied, open the ellipsis (…) menu in the Chat view → Show Agent Debug Logs

If a skill in the Skills/ folder is not being discovered by VS Code, check the following:

Every SKILL.md file must start with YAML frontmatter containing name and description fields. Without this, VS Code cannot register the skill.

---
name: my-skill-name
description: >-
  A description of what the skill does. Include USE FOR and DO NOT USE FOR
  trigger phrases to help Copilot know when to load it.
---

# Skill Title

Content starts here...

There must be a blank line between the closing --- delimiter and the first line of content. Some parsers fail to separate the metadata from the document body without it.

Correct:

---
name: my-skill
description: >-
  My skill description.
---

# My Skill

Incorrect:

---
name: my-skill
description: >-
  My skill description.
---
# My Skill

After adding or fixing a skill file, you must reload VS Code (or start a new chat session) for the skill to be discovered.

To confirm a skill is registered:

1. In the Chat view, click the gear icon (Configure Chat) — or run Chat: Open Chat Customizations from the Command Palette
2. Select the Skills tab and look for your skill in the list
3. Alternatively, type /skills in chat to open the Configure Skills menu

If a skill doesn't appear, open the ellipsis (…) menu in the Chat view and choose Show Agent Debug Logs to see why it failed to load (usually a frontmatter or name/directory mismatch).

• Reference/copilot-cli-model-routing.md — 4-tier model-routing policy for the GitHub Copilot CLI (Executors / Implementers / Tech Leads / Architects). Reference-only; not auto-attached. The document was written against the early-2026 lineup; a banner at the top maps the older model IDs (Opus 4.5 / 4.6, GPT-5.1) to the current ones (Opus 4.7, GPT-5.4 / 5.5). A full rewrite is planned post-1.1.0.

• The Agentic Operating Model — a 1h/2h/4h presentation and workshop on versioned, agent-assisted knowledge work. CopilotAtelier is used as the reference exemplar of a mature personal atelier (Module 3 — Your Atelier — Customization as Code; Module 8 — A Mature Personal Atelier) and as the cross-machine instruction-sync pattern. The workshop also publishes complementary, project-level material that pairs well with this repo:
  • Agentic Knowledge-Work Patterns — ten reusable patterns for applying the operating model beyond code.
  • Memory Bank Template — a tool-neutral starter set for per-project memory banks.

This project is licensed under the MIT License — see the LICENSE file for details.