9 protocols, sorted by the job they do. Short operating rules you can hand a teammate. Each one names what to check, who owns the call, and when to stop. Print them or paste them into your team wiki.
Free. No email required. Last updated August 15, 2026.
Protocol
The claim register
What it does. Every public claim carries a source, an owner, and a date. When to use it. Stand it up once, then update it whenever copy changes.
Make one row for every factual claim you publish. Performance numbers, capability statements, comparisons, and superlatives all count.
Fill six columns: the claim as written, where it runs, the link that proves it, who verified it, the date they verified it, and the date it needs checking again.
Add a seventh column that is easy to skip and matters most: what this claim does NOT say.
If the proof column is empty, the claim is not true yet. Change the sentence before it ships, do not reword it.
Re-check the register on a set day each quarter. Anything past its re-verify date comes down until someone re-proves it.
Checklist
CLAIM REGISTER
Claim (as written) | Where it runs | Proof (link) | Verified by | Verified on | Re-verify by | What this does NOT claim
Rules:
1. No proof link means the claim is not true yet. Change the sentence.
2. Rewording an unbacked claim does not fix it. It is the same problem in new words.
3. Every number gets a re-verify date. Numbers go stale quietly.
4. The "does not claim" column is the one that protects you in a review.
Why it works. Teams verify the claims they delete and approve the ones they write. The register catches the second kind, which is where the trouble actually is.
Protocol
The five lane pre-launch check
What it does. Five separate reviewers, each judging one thing, none covering for another. When to use it. Before anything goes live to customers.
Claims. Can every factual sentence be traced to a source someone opened today?
Render. Does it hold up on a phone, in the actual inbox or ad platform, and in dark mode? Not in the design file. In the thing.
Method. Was the QA capable of catching a problem? If the check could not have failed, it did not check anything.
Live. After launch, open the real link from outside the company. Does it load, is it the right version, is anything gated that should not be?
The checklist itself. Once a quarter, plant a known violation and see whether the review catches it.
Checklist
PRE-LAUNCH CHECK
[ ] CLAIMS: every factual sentence traced to a source I opened
[ ] RENDER: checked on a phone, in the real inbox or platform, in dark mode
[ ] METHOD: the QA could actually have caught a problem
[ ] LIVE: opened the real link from outside the company after launch
[ ] THE CHECK ITSELF: a planted mistake would be caught
One person per lane where you can. A reviewer who owns two lanes will
protect the one they care about more.
Why it works. One reviewer looking at everything sees the thing they always see. Five narrow reviewers catch five different failures.
Protocol
The AI slop scorecard
What it does. Catch machine-written copy before a customer does. When to use it. On every piece of external copy, no matter who or what wrote it.
Hard stop on dashes and emojis in body copy. Rewrite the sentence instead of punctuating around it.
Search for the forced negation pattern: not just X but Y, not only X but also Y, and the split version across two sentences. One is noise. Two in a piece is a fingerprint.
Search the banned list: leverage, synergy, cutting-edge, revolutionary, game-changer, supercharge, seamless, elevate, in today's landscape, more than just, it is important to note, when it comes to, in conclusion.
Check the reading level. Aim for grade 3 to 5. The level stays the same for every audience. A surgeon and a plumber both get short sentences.
Apply the compression test to each paragraph: what is this actually trying to say? If you can say it in one plain sentence, say that instead.
Checklist
AI SLOP SCORECARD
BLOCK
[ ] no dashes in body copy
[ ] no emojis in body copy
FLAG (2 or more means rewrite)
[ ] "not just X, it is Y" and its variants
[ ] banned words: leverage, synergy, cutting-edge, revolutionary,
game-changer, supercharge, seamless, elevate, in today's landscape,
more than just, it is important to note, when it comes to, in conclusion
CHECK
[ ] reading level grade 3 to 5
[ ] every paragraph survives the question: what is this actually saying?
Why it works. Readers cannot always name why copy feels machine made, but they feel it and they trust it less. These are the specific tells doing the damage.
Protocol
The instrument warm up
What it does. Prove the tool can speak before you believe it said nothing. When to use it. Any time a report, dashboard, or check comes back clean or zero.
Before you accept a zero, feed the tool a case that must produce a result. A test conversion, a known broken link, a deliberately wrong claim.
If nothing shows up, the tool is silent, not the subject clean. Fix the tool first.
Run it twice. A single clean run and a broken run look identical.
Be suspicious of numbers that are absurd for the thing measured. An impossible number is the instrument telling you it is wrong.
Never accept a summary of a check in place of the check. Open the thing.
Checklist
INSTRUMENT WARM UP
Before believing a zero or a clean result:
[ ] fed the tool a case that MUST produce output
[ ] the tool actually produced it
[ ] ran the check twice
[ ] the numbers are plausible for what is being measured
[ ] I looked at the artifact, not a summary of it
A silent tool and a clean subject look exactly the same from here.
Why it works. Zero conversions and zero tracking look identical on a dashboard. So do a clean brand review and a review nobody ran.
Protocol
Run, review, never
What it does. Decide what AI is allowed to do, one job at a time. When to use it. Set it up once with your team, then revisit monthly.
List the jobs your team repeats at least twice a month.
Put each on one of three lists. RUN: AI does it alone, because a wrong answer costs almost nothing. REVIEW: AI drafts, a person decides. NEVER: AI does not touch it.
For everything on NEVER, write one sentence saying why. Those sentences are your real AI policy.
Name a person, not a team, for every job on REVIEW. Marketing owns it means nobody owns it.
Move one job at a time, and keep the ability to move it back.
Checklist
RUN / REVIEW / NEVER
RUN (AI alone)
- checkable at a glance, cheap to be wrong about
REVIEW (AI drafts, a person decides)
- owner: [NAME], not a team
- what they check: [ONE LINE]
NEVER (AI does not touch)
- [job] because [one sentence]
Move one job at a time. Keep the ability to move it back.
Why it works. Two lists force a team to either trust AI completely or do everything by hand. Three lists let trust grow one job at a time.
Protocol
The five seats
What it does. Match the job to the kind of thinking it needs, and keep authority human. When to use it. When you are deciding which AI does what, or defending an AI budget.
Architect. Ambiguous work where framing matters: strategy, offers, campaign direction, concepts. Your most capable model, used sparingly.
Builder. Bounded work with a right answer: research with sources, analytics, claim checks, QA.
Production. High volume, low judgment: repurposing, formatting, captions, metadata, variants. Your cheapest capable option.
Independent reviewer. Scores work against the brief. This seat may never grade its own output.
Approval. Publishing, spending, customer actions. This seat is a person, always. No model holds it, no matter how good it gets.
Checklist
THE FIVE SEATS
ARCHITECT ambiguous, framing matters strategy, offers, concepts
BUILDER bounded, has a right answer research, analytics, QA
PRODUCTION high volume, low judgment repurposing, formatting
REVIEWER scores against the brief never grades its own work
APPROVAL publishes, spends, commits a person, always
Two rules that matter more than the table:
1. A more capable model never earns more authority.
2. The reviewer is never the writer.
Why it works. Routing by seat instead of by tool makes quality and cost the same decision, and it keeps the publish button in a human hand where it belongs.
Protocol
What your AI should know before it acts
What it does. Write down the facts once so nobody re-derives them every session. When to use it. Set up once per team, then keep it current as a habit.
Write the facts your AI needs: who the customer is, what we sell, what we sound like, what we may never claim.
Every fact carries the proof next to it. A link, a screenshot, a dashboard view. No proof means it is an opinion, so label it as one.
Every fact carries a date and a expiry. Campaign results age in weeks. Positioning ages in months. A stale fact is not false, it is unverified, and it gets flagged not quoted.
Every open item carries a done condition someone else could check, written as a test and not an intention.
Every blocked item names the person it is blocked on. Waiting on nobody is how work disappears.
Decisions get appended, never edited. Reversing one is a new entry that says it replaces the old one, so the same argument is not had twice.
Checklist
TEAM MEMORY
FACTS
fact | proof (link) | verified on | re-check by
OPEN
item | done when [someone else could check this] | owner
BLOCKED
item | blocked on [a person, by name] | since
DECISIONS
decision | why | date | status: standing / replaced by [entry]
Rule: if you cannot name the proof, it is not a fact. Write it as an opinion
or go find the proof.
Why it works. Most teams pay for the same lookup every week because the answer lives in someone's chat history. Write it once, prove it, date it.
Protocol
The outsider seat
What it does. Put one person with no context in front of the work before it ships. When to use it. On naming, positioning, pricing pages, and anything you have looked at too long.
Find someone who has not been in the meetings. Any department. The newer the better.
Show them only what a stranger would see. No setup, no preamble, no explaining what you meant.
Ask five questions: what is this, who is it for, what are you being asked to do, where did you get confused, what would you guess it costs.
Write down the order they got confused in. That order is your fix list.
Do not defend anything. The moment you explain, the test is over.
Checklist
THE OUTSIDER TEST (15 minutes)
Show them only what a stranger sees. Then ask:
1. What is this?
2. Who is it for?
3. What are you being asked to do?
4. Where did you get confused? (write the ORDER)
5. What would you guess it costs?
Rule: you may not explain. The first explanation ends the test.
Why it works. Everyone who built the thing knows what it is supposed to say, which is the exact reason they cannot see what it actually says.
Protocol
The owner gate
What it does. Name the human who owns the call before the work starts. When to use it. On any AI job that touches a customer, money, or the public.
Write the job in one sentence, small enough to judge.
Name one owner. One person, by name.
Name what the reviewer checks, in plain words. Facts, numbers, privacy, tone, or customer readiness.
Write the stop rule before the first run. What would make you pull it back?
Keep the AI away from the send button. Drafting, sorting, and preparing are fine. Sending, publishing, and spending stay human.
Checklist
OWNER GATE
Job: [one sentence]
Owner: [one person, by name]
Reviewer checks: [facts / numbers / privacy / tone / ready for a customer]
Stop rule: we pull this back if [one sentence]
The AI may not: send, publish, spend, or take any action we cannot undo
Why it works. Teams adopt an AI job exactly as far as the question who is accountable when it is wrong has an answer, and not one step further.
Go deeper
Once the rules hold, you can hand a job to an agent.
A blueprint is a job description for an AI worker. It names the job, the guardrails, and the human who still owns the decision.