If AI work only happens in private chats, you do not have a practice. You have hobbies that sometimes reach customers.
This is a 30 to 45 minute weekly loop for one owner and, optionally, one buddy. It is not a steering committee. If it needs slides, you have already lost the week.
Aside. A weekly AI meeting that does not look at a single real send is a book club. Book clubs are fine. They are not operations.
The clock
Set a recurring 40-minute block. Same day. Same doc. If you skip two weeks, restart. Do not "catch up" with a 90-minute archaeology session.
Suggested split:
- 10 min — misses and near-sends
- 10 min — packets and library
- 10 min — tools and access
- 10 min — one experiment or one kill
Minute 0–10: Misses
Open the log. If you have no log, that is this week's experiment: a three-column table in a shared doc (date, job, what we almost shipped).
Look at anything that:
- Failed the customer test
- Invented a fact
- Was rewritten from scratch
- Made someone ask "why did we use a model for this"
For each miss: was it packet, source, hygiene, or "should have been a no"? Change one artifact. Do not add a new tool.
Minute 10–20: Promote or kill
List the repeating packets (prompts, reply cards, SOP jobs). For each:
- Promote: a second person used it and it passed eval. Put it in the shared folder with a date.
- Keep: still personal, still useful. Leave it, but it is not "the team's."
- Kill: unused for three weeks, or it fails eval. Delete from the shared folder. Chat history can rot on its own.
Stale shared prompts are more dangerous than no prompts. People trust the folder.
Minute 20–30: Doors and data
Quick pass, spoken out loud:
- Who joined, who left, whose API key is still in a spreadsheet
- Did anyone paste a customer dump this week
- Are we still on consumer ChatGPT / Claude / Gemini for work that should be a team plan
- Any model or product change that breaks a packet (a default model swap is a change)
If you cannot answer "which door," see the paste-hygiene lesson and fix that before new use cases.
Minute 30–40: One experiment or one kill
Pick one:
- Add one reply-library card from a real thread
- Dry-run a packet on two new cases
- Move one job from consumer to a contracted door
- Kill a use that failed the when-not-to-use-ai rule
- Time a job with and without the model (clock on the phone, not a story)
Write the result in the same doc in five lines: what, result, keep/kill, owner, date. No dashboard required.
Copy-paste: agenda the model can prep (you still meet)
Using only the LOG I paste, draft a 40-minute agenda.
>
Rules: do not invent misses. If the log is thin, say so and propose what to collect next week. Suggest one promote, one kill, and one experiment, each with a reason from the LOG. Forty minutes includes no slide time.
>
LOG:
[bullets from the week]
Roles
- Owner (one name): keeps the doc, ends the meeting on time, records keep/kill.
- Buddy (optional): brings one miss and one "this was actually good."
- Everyone else: reads the doc. They do not need to attend.
When the owner is out, the buddy runs a 20-minute version or you skip. You do not rotate in a stranger to "keep the ritual alive."
What this is not
- Not an innovation lab
- Not a tool bake-off every Friday
- Not a mandate to increase "AI usage" as a KPI
- Not a place to demo a new agent that has never had a human gate
Usage as a KPI produces paste theater. Track misses caught and time on repeating jobs instead, if you track anything.
Failure modes
- Museum folder: 40 prompts, two in use, none dated.
- Shame culture: people hide misses, so the log is empty and the customers see the misses instead.
- Tool of the week: the experiment is always a new subscription.
- Ownerless: "the team" owns it. Nobody opens the doc.
Download
Weekly agenda and log: /downloads/weekly-ai-ops-cadence.md
You are done when
The doc has three weeks in a row with at least one miss or one kill. Clean weeks with only wins are a smell.