Plug AI4BCM Guidance into the assistant you already use.

The router is one folder, ask-ai4bcm/, holding SKILL.md and agents/openai.yaml. It follows the Agent Skills format (https://agentskills.io/specification, read 2026-09-10), so a client that reads that format reads this one.

Edit ask-ai4bcm/, never skills/ask-ai4bcm/. The second one is generated. A portable plugin package finds skills only under skills/<name>/, and a skill carries its own reference files, so tools/build_chatgpt_package.py writes a copy of the router there with units/ inside it and the router's links repointed to match. Regenerate it after any change to ask-ai4bcm/ or units/; tools/build_chatgpt_package.py --check fails if the two have drifted apart.

Three words this page uses

You do not need them for the guidance itself, only for installing it. They are here rather than in units/glossary.md, which is a BCM glossary.

• Plugin — a packaged folder a client installs in one step, so you do not copy files by hand. This repository is packaged as one.
• Marketplace — the place a client fetches plugins from. It need not be a shop: here it is this GitHub repository, which lists itself, and nothing is bought.
• MCP — Model Context Protocol, a way for a client to reach a live server while it answers. The router uses none. It is named here only because the ChatGPT page explains why the package deliberately declares no MCP server.

Three routes, and what each one brings

As a Claude Code plugin. Two commands, and the units come with it, so every link in the router resolves:

/plugin marketplace add AI4BCM/guidance
/plugin install ai4bcm-guidance@ai4bcm

With the skills CLI, which installs into whichever agent directories it finds:

npx skills@latest add AI4BCM/guidance

Update later with npx skills update. This copies the skill folder alone. The router then names units that are not on disk beside it, and says so; read them from the repository, or clone it and use the folder route instead.

By hand, as a folder. Clone the repository and copy ask-ai4bcm/ and units/ together into the directory your client scans. One page per client, each written from that client's own current documentation with the URL and the date it was read:

• Claude Code
• Codex
• GitHub Copilot
• ChatGPT — the plugin route needs a Business, Enterprise, Healthcare or Edu workspace

What the router does, and what holds it back

It routes and nothing else. It names the unit to read, the prompt to run and the maturity level the work needs, then hands over. Running the prompt is yours.

Two of the four clients can be told to leave the skill alone until you ask for it by name. Claude Code reads disable-model-invocation: true in SKILL.md; Codex and the ChatGPT desktop app read policy.allow_implicit_invocation: false in agents/openai.yaml. GitHub Copilot reads neither, and its own documentation says it loads a skill when it judges it relevant.

Source: https://code.claude.com/docs/en/skills, read 2026-09-10.

The short way, as a plugin

This repository is its own plugin marketplace. Two commands, and the units are installed alongside the router, so its links resolve:

/plugin marketplace add AI4BCM/guidance
/plugin install ai4bcm-guidance@ai4bcm

/plugin then lists it as ai4bcm-guidance@ai4bcm, carrying one skill, ask-ai4bcm. Updates come through /plugin update ai4bcm-guidance.

Where the folder goes

Personal, available in every project on your machine:

~/.claude/skills/ask-ai4bcm/SKILL.md

Inside one repository, available to anyone who clones it:

.claude/skills/ask-ai4bcm/SKILL.md

The command name comes from the directory name, so the directory must stay ask-ai4bcm.

Steps

1. Clone or download this repository.
2. Copy ask-ai4bcm/ into ~/.claude/skills/ or into .claude/skills/ in your project.
3. Copy units/ alongside it, or clone the whole repository and symlink ask-ai4bcm/, so the relative links from SKILL.md into ../units/ resolve.
4. Type /ask-ai4bcm in a Claude Code session.

What the frontmatter does here

SKILL.md carries disable-model-invocation: true. The documentation states the effect. You can still invoke the skill with /ask-ai4bcm, the model cannot load it on its own, and its description stays out of the model's context until you ask. That is the intended contract for this router.

Source: https://learn.chatgpt.com/docs/build-skills, read 2026-09-10, reached by permanent redirect from https://developers.openai.com/codex/skills/.

Where the folder goes

Codex reads skills from repository, user, admin and system locations. The documented directories:

$CWD/.agents/skills
$CWD/../.agents/skills
$REPO_ROOT/.agents/skills
$HOME/.agents/skills
/etc/codex/skills

For a router you want everywhere, $HOME/.agents/skills/ask-ai4bcm/ is the one to use.

The short way

npx skills@latest add AI4BCM/guidance

The skills CLI finds ask-ai4bcm in the repository and copies it into the agent directories it detects, recording what it installed in skills-lock.json; npx skills update refreshes it. It copies the skill folder alone, so the units named in the router are not beside it. Either read them from the repository, or take the folder route below, which puts units/ where the router expects.

Steps

1. Clone or download this repository.
2. Copy ask-ai4bcm/ into ~/.agents/skills/.
3. Copy units/ alongside it, or clone the whole repository and point ~/.agents/skills/ask-ai4bcm at the clone, so the relative links into ../units/ resolve.
4. In the Codex CLI or IDE extension, type $ask-ai4bcm. /skills lists what Codex has found.

Codex detects skill changes without a restart.

What agents/openai.yaml does here

The file beside SKILL.md reads:

policy:
  allow_implicit_invocation: false

With that setting, only explicit invocation works. It is how Codex is told what disable-model-invocation: true tells Claude Code.

Sources, each with the UTC time it was read:

• Skills in ChatGPT, 2026-09-10 10:51
• Plugins in ChatGPT and Codex, 2026-09-10 10:52
• Creating and editing GPTs, 2026-09-10 10:53
• Skills & Plugins, 2026-09-10 10:50
• Build skills, 2026-09-10 10:50
• Skill controls, 2026-09-10 10:50

Re-read 2026-09-10 for the plugin package, with the sources that decided its shape:

• Skills in ChatGPT, 17:49
• Plugins in ChatGPT and Codex, 17:51
• Importing and syncing plugin marketplaces from GitHub, 17:52
• Package your plugin, 17:48
• Submit your Claude Code plugin to OpenAI, 17:48
• Agent Skills specification, 17:50

The help.openai.com articles were read in a browser; they answer HTTP 403 to curl from some networks, and a check reporting them unreachable has hit that 403 rather than a dead page. The learn.chatgpt.com, developers.openai.com and agentskills.io pages serve Markdown at a .md suffix and answer plain curl.

Who can install this in ChatGPT

Skills are the thing ChatGPT loads, and they are gated by plan. The help centre states it in one line: "Skills are available to eligible ChatGPT Business, Enterprise, Healthcare, and Edu users, subject to workspace settings and product availability." Even inside one of those workspaces an administrator can turn skills off for your role, so check with whoever runs the workspace before you spend time on this.

On a personal account — Free, Go, Plus or Pro — no route to the router exists on any ChatGPT surface. The obvious workaround is closed too: new GPT creation and publishing are no longer available on personal accounts, so you cannot wrap the guidance in a custom GPT either.

Use the guidance without the router instead; what it costs you is a lookup. units/ is prose meant to be read. Open the unit for your stage, or start from units/levels.md under Starting guide, and take the prompt you need from units/prompts/. Classify your material against units/data-rules.md first, which is the one step the router would have insisted on before anything else. The print edition of the guidance is written for exactly this reader and carries the same material in reading order.

What works today, through Codex

The documented filesystem locations for skills belong to Codex, which reads $HOME/.agents/skills among others, and Codex runs inside the ChatGPT desktop app as well as from the CLI. So on a machine where both are installed, follow codex.md, put the folder in ~/.agents/skills/ask-ai4bcm/, and @ask-ai4bcm selects it. The ChatGPT desktop app lists skills in its sidebar.

This is a developer action. It needs a checkout of the repository, a terminal and a machine you control, and it reaches that one machine and nobody else in your team. Workspace skill features stay subject to the plan gate above whatever sits on your disk.

The plugin route

A plugin can bundle skills, and plugins work in Chat and Work across ChatGPT on the web, desktop and mobile. That is the route worth having, because it reaches the phone and the browser as well as the laptop, and a workspace administrator can hand it to everyone at once rather than to one machine at a time.

The package exists now. It carries the router and the units together, which is the whole point: the router names units by relative link, and a router whose links dangle is worse for you than no router at all. The units travel inside the skill folder, at skills/ask-ai4bcm/units/, so the links resolve wherever the folder lands.

It declares no MCP server. MCP, the Model Context Protocol, is a way for a client to reach a live server while it answers a question; the router needs none, because it routes you to files rather than fetching anything. Declaring one anyway would cost you the web: a plugin that declares an MCP server is labelled Desktop only and cannot run on ChatGPT on the web, even when the server sits behind a remote HTTPS URL. That is the one change that would silently take away the surface this package exists to reach, so the package contains no mcp.json and no .mcp.json.

Route A: import the marketplace from GitHub

For a workspace administrator who wants the guidance to stay current. ChatGPT re-reads the repository daily.

1. Go to Workspace settings > Plugins, select Add, then Import marketplace.
2. In Source, enter https://github.com/AI4BCM/guidance. Enter the repository URL only, with no branch and no folder path after it.
3. Leave Path empty. The catalogue this repository publishes is at .claude-plugin/marketplace.json in its root, which is one of the manifest names the importer accepts.
4. Leave Branch, tag, or commit empty to follow the default branch, or name a tag to pin the wording your organisation reviewed. A pinned commit stays where you put it and never syncs forward.
5. Select Import marketplace and authorise GitHub access when prompted.
6. Open the imported ai4bcm-guidance plugin and set its Installation policy: Available lets eligible members install it themselves, Installed puts it on their accounts for them.

Import does not read the policy values written in the repository, so step 6 is not optional — until you set a policy, nobody has the plugin. To pull a change before the next daily sync, open Marketplaces, select the marketplace and choose Sync now.

The GitHub account you import with needs read access to the repository, and future syncs keep using that account. If that person leaves, a new administrator imports the same source again with their own GitHub connection rather than repairing the old one.

Route B: upload the skill folder

For a workspace with no GitHub connection, or for one person testing the router before the organisation commits to it. This installs the skill alone rather than the plugin.

Take skills/ask-ai4bcm/ from the repository — the whole folder, including the units/ directory inside it — and upload it under Plugins > Skills > Create > Upload from your computer. An administrator can do the same from the admin Skills page with + Add skill, which puts it in front of the workspace instead of one account.

ChatGPT scans an uploaded skill before it becomes available, and may mark it Needs Review before you can use it. Review it yourself as well: it is a folder of instructions and text, it came from outside your organisation, and the scan is not a substitute for your own judgement about what your people are told to do with BCM material.

What is not being claimed here

Neither route has been run in a real Business, Enterprise, Healthcare or Edu workspace, because this guidance was packaged on a machine that has no such workspace. What has been checked is the package: the archive was built, extracted, and every one of the router's 46 relative links resolved from inside it. Whether ChatGPT accepts the upload, and what its scan says about a router that asks not to be invoked on its own, are answered the first time an administrator with a workspace tries it. If you are that person and something here is wrong, the repository takes issues.

Sources: https://docs.github.com/en/copilot/concepts/agents/about-agent-skills and https://docs.github.com/en/copilot/how-tos/copilot-cli/customize-copilot/add-skills, both read 2026-09-10.

Where the folder goes

Inside a repository, any of:

.github/skills
.claude/skills
.agents/skills

Personal, shared across your projects:

~/.copilot/skills
~/.agents/skills

The short way

npx skills@latest add AI4BCM/guidance

The skills CLI finds ask-ai4bcm in the repository and copies it into the agent directories it detects, recording what it installed in skills-lock.json; npx skills update refreshes it. It copies the skill folder alone, so the units named in the router are not beside it. Either read them from the repository, or take the folder route below, which puts units/ where the router expects.

Steps

1. Clone or download this repository.
2. Copy ask-ai4bcm/ into ~/.copilot/skills/, or into .github/skills/ in the repository you work in.
3. Copy units/ alongside it, so the relative links from SKILL.md into ../units/ resolve.
4. In the GitHub Copilot CLI, invoke it by name: /ask-ai4bcm.

gh skill in the GitHub CLI searches for, installs, updates and publishes agent skills, and copilot skill add <FILE | URL | DIRECTORY> adds one from a path.

Copilot reads the skill in the cloud agent, in Copilot code review, in the GitHub Copilot CLI, in the GitHub Copilot app, and in agent mode in Visual Studio Code and JetBrains IDEs.

What Copilot does not read

Copilot reads neither disable-model-invocation nor policy.allow_implicit_invocation. Its documentation says Copilot loads a skill when it judges the skill relevant to the task, and that skills in .github/skills are also available to Copilot code review. So in Copilot the router can open on its own, which it cannot do in Claude Code or Codex. It still routes and nothing more, naming a unit, a prompt and a level, and running nothing itself.

Every section above is rendered at build time from install/ in the AI4BCM Guidance repository, so this page cannot drift from the document it publishes. Each tool page names the vendor documentation it was written from and the date it was read.