The protocols — how an AI employer actually runs the floor

Prompts are not a company. Protocols are. Below is the operating system we use when multiple AI employees share a mission — distilled from scars, then enforced until they hold under stress.

If you only remember one line: structure beats intention.

1. BST — talk to the owner

Briefest · Simplest · Transparent.

When reporting to the human: short true answer first. When handing a deliverable (prompt, email, post, code): complete unless asked otherwise. Cutting a tool prompt “to be brief” is a protocol violation — that starves the machine.

2. Announce → do → report

Silent background work in a multi-seat room causes collisions. Every non-trivial action:

1. One line: what I’m about to do

2. Do it

3. One line: what shipped, with path or proof

See also: Compose, re-read, send for the external-send variant.

3. One owner per artifact

First claimer owns the path. Second seat reviews — never rebuilds in parallel. Dual packs are how nights die. Full scar: One owner per artifact.

4. Read before speak (multi-agent)

Before posting in a shared channel or council log: re-read peers and the human’s latest north-star note. The per-seat cursor is truth — not a shared size file, not memory of “I think I saw that.”

Evolution of this rule: How we employ multiple AIs.

5. Dual review before public

Nothing public ships on one seat’s say-so:

• Scrub — secrets, client exposure, false claims
• Premium — brand tokens, voice, CTA honesty

Silence is not approval. A missing reviewer does not unlock new public content.

6. If a rule isn’t code, it’s a wish

Protocols that only live in chat will be forgotten after compaction. Encode them: claim registries, send gates, deploy checklists, scorecard honeypots. Scar: If a rule isn’t code.

7. Never claim fixed until verified

“Shipped” ≠ “fixed.” Independent check first — live URL, real form post, second pair of eyes. Green dashboards can still be wrong: Green lights lie.

8. Order of craft

Plan → write → judge → render → publish.

Never pretty-render a thin skeleton twice and call it a legacy.

How the protocols fit the product

The public brand teaches AI employment. Internally we are the case study: job descriptions (lanes), owned files (claims), coded guardrails (CLI/gates), escalation (human chair), performance reviews (scars → rules).

That loop is the business. The website and the blog exist to make it transferable.

Next: Employing AI vs using AI · Multi-agent employment · Score your operation