One specialist written down before a single thing is built. Job, access, output, off switch, and a test that can fail. If you cannot complete it, you have learned the agent is not ready, which is worth more than a half-built one.
Use it on the one seat your org chart marked first, after the coordinator spec exists. Use it again for each later seat, one at a time.
Do not use it on a box that failed the short-and-testable check. Go back and split the box first.
One box from your chart, and honesty about what it touches. Which files, which accounts, which customers. If you do not know what it needs access to, that is the first finding, not a detail to fill in later.
You are helping me define ONE AI agent before I build it. Your job
is to be strict. If I cannot answer something, say so plainly and
stop rather than filling the gap with a reasonable guess.
Do not write code. Do not recommend tools or platforms.
Ask me about these five fields, ONE AT A TIME, and push back if an
answer is vague:
1. JOB. What single thing does it do, stated so narrowly that I
could tell whether it did it or not? If my answer contains the
word "and", it is probably two agents.
2. ACCESS. Exactly what can it read and write, and nothing more.
Ask me what it must NOT reach.
3. OUTPUT. Where does its work land, in what format, and who looks
at it before anything happens because of it?
4. OFF SWITCH. How do I stop it, how fast, and what breaks when I
do?
5. TEST. A check that can actually FAIL. Not "does this look
right." A number that clears a bar, or a known answer it has to
reproduce. If I offer you a subjective check, reject it and ask
again.
Then write the job description as a single page under those five
headings.
Finally, do one of two things. If all five are solid, say the agent
is ready to build and name the one field most likely to be wrong in
practice. If any field is weak or blank, say clearly that the agent
is NOT ready, name the missing field, and tell me what to go find
out. Do not soften that verdict.
Run it per agent, never in batches. Ten minutes. Expect roughly a third of your seats to come back not ready on the first pass, and treat that as the prompt working.
Keep the finished page next to the agent. When it misbehaves later, the TEST field tells you whether it broke or whether you never defined what working meant.
A test with a number in it. An access list with a stated boundary on what it must not reach. A named human between the output and anything consequential. And a verdict at the end that is willing to be no.
A job containing the word "and", which is two agents wearing one coat. Granting broad access because narrowing it is tedious. Accepting "I will review it" as the test, which is not a test because it cannot fail on its own. Building anyway after a not-ready verdict.
The field people skip is the off switch, and it is the one that decides whether you will actually let the thing run. An agent you cannot stop quickly is an agent you will supervise constantly, which costs more than doing the work yourself.
Source: Build to Thrive, The AI Blueprint, week of August 17, 2026.