Shared: Memo B: how we pick a winner for Cletus
Memo B: how we pick a winner for Cletus shared by Claude (diligence team). Discuss it on the Discussion page.
Fri, Oct 16
Joe, Reid, Keegan · How to use this site
First call chaired by Claude. Claude joined partway and filed the start from Joe's summary. Decided: focus on one region at a time, and building Cletus, the AI co-worker, is the experiment. Keegan brings the Ector County records on Monday.
What this call produced 15
File: Memo B: how we pick a winner for Cletus
Not filed yet
File: Memo A: which AI-with-its-own-desk versions to test
Not filed yet
File: Google Meet notes for the whole call (Gemini)
Not filed yet
To-do T25: Talk about how to have a real back-and-forth with Claude during a call, alone or as a team, like we get in the terminal
Filed under Meet Notes app
To-do T24: Find the cheapest, most certain way to turn old handwritten county indexes into a usable index, starting with Martin County
Filed under County data plan
To-do T23: Build the check that measures whether each task got done right, done by code, not by hand
To-do T22: Choose the agent setup for the experiment from the two research write-ups
Filed under Experiment 1: four ways to build the co-worker
To-do T21: Finish the two tests and send Joe a write-up: what was done, why, how, and the results, saved in a repo
Filed under Experiment 1: four ways to build the co-worker
Decision D4: Record everything said in these calls, even ideas we don't use yet, so old ideas can come back when they fit
Filed under Meet Notes app
Decision D3: Building the AI co-worker, Cletus, is the experiment: it routes documents, runs title and handles tasks, and we keep building until we are satisfied
Filed under Experiment 1: four ways to build the co-worker
5 older
Decision D2: Build the AI co-worker, named Cletus, to route documents, run title and handle tasks
Filed under Experiment 1: four ways to build the co-worker
Decision D1: Focus on one region at a time, not many counties at once
Filed under County data plan, AI server hardware
To-do T20: Keep running the early conflict checks and settle the experiment's benchmarks
Filed under Experiment 1: four ways to build the co-worker
To-do T19: Bring the Ector County records
Filed under County data plan
To-do T18: Find the Google Meet transcript of today's call and give it to Claude, so the part before Claude joined gets filed
Filed under Meet Notes app
Goals
Can the team agree how we will use the meeting app, so it helps us build the experiment and the AI workflow?
Reid and Joe agreed to run calls with goals up front and let Claude record everything; details get tuned call by call
Done when: Joe, Reid and Keegan agree on how they will use it in the next call · Housekeeping
Full close-out
Not decided yet
- Record everything said in these calls, even ideas we don't use yet, because Nobody said yes before the call ended. Waiting on: A yes at the next call
- When the early conflict checks and benchmarks are due, because No date was set. Waiting on: Joe, Reid and Keegan
To-dos by person
- No owner yet: Find the Google Meet transcript of today's call and give it to Claude, so the part before Claude joined gets filed
- No owner yet: Build the check that measures whether each task got done right, done by code, not by hand
- No owner yet: Find the cheapest, most certain way to turn old handwritten county indexes into a usable index, starting with Martin County
- No owner yet: Talk about how to have a real back-and-forth with Claude during a call, alone or as a team, like we get in the terminal, by next call
- Joe (owner unconfirmed): Choose the agent setup for the experiment from the two research write-ups, by this weekend
- Joe, Reid, Keegan (owner unconfirmed): Keep running the early conflict checks and settle the experiment's benchmarks
- Keegan (owner unconfirmed): Bring the Ector County records, by Monday
- Reid (owner unconfirmed): Finish the two tests and send Joe a write-up: what was done, why, how, and the results, saved in a repo
Topics carried forward
- Scanner and storage
- AI server hardware
- The answers we test against
Nothing in: Which part of the vision moved, Project stage changes, Drafts waiting for review.
Claude's replies · 15
Shared: Memo A: which AI-with-its-own-desk versions to test
Memo A: which AI-with-its-own-desk versions to test shared by Claude (diligence team). Discuss it on the Discussion page.
Decision confirmed: Cletus reaches our files directly
Joe confirmed (Oct 18): files must be reachable by Cletus directly. Wherever our files live (SharePoint today), Cletus must be able to open, save and move them through that system's official connection for software, with no fragile workarounds.
Decisions confirmed
Joe confirmed (Oct 18)
- Decided: size the AI hardware with a 25% buffer above the forecast workload.
- Decided: the experiment's control is "no AI", plain code matching check stubs and division orders. Every AI version has to beat it. This replaces the Oct 13 plan of using the first AI version as the baseline.
Still open
- Files must be reachable by Cletus directly. Wherever our files live (SharePoint today), Cletus must be able to open, save and move them through that system's official connection for software, with no fragile workarounds. Reid already does this with SharePoint. Waiting on Joe's yes.
Google Meet's notes for the whole call
Google Meet's own notes for the whole call are now attached (10:03–11:23, including the first hour, before Claude joined): Google Meet notes (Gemini). Here's what they add to this record.
Decisions to confirm (not yet on the board)
- Size the AI hardware with a 25% buffer above the forecast workload (Joe and Keegan).
- Whatever document system we use must move files smoothly through its API, with no brittle workarounds (Dropbox, Box and SharePoint were discussed).
- The experiment's control is "no AI": plain code matching check stubs and division orders, from Reid's early test. This conflicts with Oct 13 ("the first AI version is the baseline"), so confirm which one stands.
New facts
- Reid's source-conflict audit compared our answers from HubSpot with the Excel inventory across 40–50 deals, to settle what we score against.
- Mail scanning: Kelsey bulk-scans with blank separator pages and Claude names the files. 49 documents went from 3.5–4 hours to 1 hour; later batches took 15 minutes.
- County data prices: Ector County $2,500; Lea County (NM) quoted $250,000. That's why we focus on one region and read the text with OCR instead of breaking down every document.
- Reid's reading pipeline already exists: regular OCR first, then Gemini or Claude vision for poor pages. It costs per page until we have our own server.
Owners from Google's notes (to confirm)
| Owner | To-do |
|---|---|
| Reid | Build the code check that measures whether each task was done right (T23) |
| Reid | Write up the two tests, push them to a repo, and email the benchmark results (T21, T20) |
| Joe | Pick the agent setup from the research (T22), and draft the metrics for judging Cletus |
| Joe | Share the presentation page so everyone can open it, and push the code fixes after review |
| Keegan | Bring the Ector County records by Monday (T19) |
| Group | Confirm the answer keys are in place before the experiment |
| Group | Put the notes app on a live server, keep shared files in one place, and show the terminal in the app (T25) |
T18 (get the Google Meet notes) is done with this post.
Shared: Google Meet notes for the whole call (Gemini)
Google Meet notes for the whole call (Gemini) shared by Google Meet. Discuss it on the Discussion page.
Open questions
Added from the end of the call
- Keegan's update: his contact will send Ector and Midland County records, likely Monday or Tuesday. Older indexes are handwritten, and turning them into a usable index is the base for running title. Reading them with Claude's vision works but is expensive.
- To-do, Reid: finish the two tests and send Joe a write-up (what, why, how, results), saved in a repo.
- To-do, Joe: choose the agent setup for the experiment this weekend.
- To-do, no owner yet: build the check that measures whether each task got done right, done by code.
- To-do, no owner yet: find the cheapest, most certain way to turn handwritten county indexes into a usable index, starting with Martin County.
- Note: Joe's reader already sends only the hard pages to Gemini and Claude. A local server would do that without paying per use.
Open questions
- Yes to recording everything said in these calls?
- Who owns the measuring check? (Joe suggested Reid.)
- Who owns the handwritten index test?
- When are the early conflict checks and benchmarks due?
To close: answer these now or at the next call. They show first next time either way.
Wrapping up: read-back
Goal met: the team agreed to run calls with goals up front and let Claude record everything. Details get tuned call by call.
Read-back: say "Hey Claude, yes to all" or fix any
- Record everything said in these calls, even ideas we don't use yet, so old ideas can come back when they fit.
- Keegan brings the Ector County records on Monday.
- Joe, Reid and Keegan keep running the early conflict checks and settle the experiment's benchmarks. When is it due?
Recap, and how meetings run
Recap of today
- Decided: focus on one region at a time, not many counties at once.
- Decided: building Cletus, the AI co-worker, is the experiment. It routes documents, runs title and handles tasks, and we keep building until we are satisfied.
- To confirm: record everything said in these calls, even ideas we don't use yet, so a good idea can come back when the time is right (Reid's point).
- To-dos: Keegan brings the Ector County records on Monday. Joe, Reid and Keegan keep running the early conflict checks and settle the experiment's benchmarks.
- Worked: blank pages between documents let one scan of a mixed stack get sorted by Claude.
How Claude keeps a meeting on track
- Start: who is here, how long, and up to three goals, each with "done when".
- During: Claude files every decision and to-do under its project or topic as it is said. If talk drifts, side ideas go to a "parked" list instead of being lost.
- Nudges: at halfway and 10 minutes left, Claude checks the goals. If one is stuck, a quiet prompt shows on the page (at most four a call). Claude never interrupts out loud.
- End: Claude reads back each decision and to-do for a yes, scores each goal (met or not), and lists what carries to next time.
- Next call: it opens with what is still open, oldest first, so nothing falls through the cracks.
Claude only keeps you on track. The group decides. Nothing counts until someone says yes out loud.
For Joe
The one goal for today is still not started: agree on how Joe, Reid and Keegan will use the meeting app so it helps build the experiment and the AI workflow. A good next step is to talk that through now. The call has three topics on its list: the scanner and storage, the AI server, and the test answers. Some open items fit those topics. Reid could approve or swap the 27 test documents (T1), someone could hand-check the 2 leases and 2 tax documents (T6), and Joe could ask Dell for the server quotes (T11). Keegan is bringing the Ector County records on Monday.
Filed from the start of the call
The start of the call is now filed, from the summary Joe pasted.
To-dos
- Keegan brings the Ector County records on Monday. Under County data.
- Keep running the early conflict checks and settle the experiment's benchmarks: Joe, Reid and Keegan. Under Experiment 1.
Decisions to confirm (say yes or fix them)
- Focus on one region at a time, not many counties at once.
- Build the AI co-worker, named Cletus, to route documents, run title and handle tasks.
Notes added
- Scanner: blank pages between documents worked, and Claude sorted the stack.
- AI server: sized for one region first.
- Test answers: early tests compare HubSpot with the Excel inventory.
For Joe
The chime is a soft sound that plays only in the runner's view, which is Joe's screen. It plays when Claude needs an answer from the group. Nobody else hears it, and it plays once for each new prompt. It's optional and starts off. A "Chime is off - turn on" button on that page switches it on for that browser.
Set up for today
Got it. Joe, Reid and Keegan are here, with about 15 minutes left.
Today's goal: agree how the team will use this app, so it helps build the experiment and the AI workflow.
The earlier part of the call: Claude did not hear it. After the call, Joe gets the Google Meet transcript (if recording or transcripts were on, it is in Joe's Google Drive, in a "Meet Recordings" folder). Claude files what was decided from it as suggestions for the next call. That is now a to-do under the Meet Notes app.
What the "chime" is: a soft sound that plays only on Joe's screen when Claude needs an answer from the group. Nobody else hears it. It is optional.
How to use it: say "Hey Claude", pause, then ask. Decisions and to-dos show up on this page and under each project on the home page.
Getting started
Claude joined partway through this call, so it did not hear what was said before now.
- Who is here?
- How long do we have left?
- What have we covered so far? A quick two-minute recap: decisions made, who owns what, and what is still open. Claude will read each one back for a yes before it counts.
- What do we want to walk out with?
Say "Hey Claude", then pause, to ask for anything.
Listening
Claude is listening. Say "Hey Claude, ..." in the call and the reply appears here.