AI Pathway build standards, as Agent Skills

Open specifications for the back-office workflows that leak money in small businesses, packaged so an AI coding agent can install one and build to it.

Free to implement under CC BY 4.0, commercially included. The only thing asked in return is the citation, because a workflow built to a shared standard can be audited, handed over, and trusted by whoever holds the seat next.

Install

/plugin marketplace add digitalgym/aipathway-standards
/plugin install build-standards@aipathway-standards

Or copy any skills/<name>/ folder into .claude/skills/ in your own project.

On ChatGPT or Codex, the same skills install as an OpenAI plugin:

codex plugin marketplace add digitalgym/aipathway-standards
codex plugin add build-standards@aipathway-standards

In the ChatGPT desktop app, add the marketplace the same way, then install build-standards from the Plugins Directory under AI Pathway.

Not on Claude? The same rules are in AGENTS.md (Codex, Copilot CLI and anything that reads the convention), .cursor/rules/aipathway-standards.mdc (Cursor) and .github/copilot-instructions.md (GitHub Copilot). All three are generated from the same list as the skills, so they say the same thing.

Running the checks

Every standard also publishes its checks as data, and packages/conformance is the harness that runs them:

npx @aipathway/conformance init booked-after-hours-build-standard
cd booked-after-hours-build-standard && npm install && node run.mjs
npx @aipathway/conformance connect     # a key for the hosted step; the owner approves in a browser

It ships a recording stub and a runner, so a build can be driven through the ports and asserted without a live phone number. The checks are not bundled into the package: the CLI fetches them from https://aipathway.com.au, so a standard corrected on Tuesday needs no release and nobody can be conformant against a stale copy. The package is the harness; the site is the standard.

The Mesh Directory

mesh/ is a directory of Grok Bots, and agents like them, ported to these standards so they work as one team: one id from the system of record on every hop, each hop woken by a named event, an open exception stopping the chain. A bot is listed by pull request with a green conformance run, and every port credits the bot it came from. See mesh/README.md and https://aipathway.com.au/explore-ai/mesh-directory.

The standards

What a standard contains

Each one carries the same six things, and a published audit holds every standard to all of them:

• The checks, in order. The order is the claim, not a presentation choice.
• The objects, with their fields. The records the workflow turns on.
• A conformance checklist, so a DIY build or a vendor can be held to it.
• A pass test a stub cannot fake.
• A provider block, naming what to stub and what not to hand-roll. Where no product covers that step it says production: none rather than staying quiet.
• A boundary: what stays with a person, and why.

These are specifications, not marketing

The standards are published because a workflow nobody can audit is worth less than one anybody can. They name what should not be automated as plainly as what should, they carry no results claims, and where the honest answer is "there is no product for this, build it yourself" they say that.

Canonical home, always current: https://aipathway.com.au/explore-ai Machine-readable: https://aipathway.com.au/llms.txt · https://aipathway.com.au/llms-full.txt · https://aipathway.com.au/AGENTS.md

Generated, not hand-written

Every file here is generated from the live standards by scripts/build-skills.mjs in the site repo, so a skill cannot drift from the specification it claims to carry. Send corrections to the standard, not to this repo: hello@aipathway.com.au.