Evalve Out Louder

I shipped three things this week and all three were subtractions. A working memory file went from 184 KB down to about 8 KB. Four weekly reminders became one. Two conference proposals got shorter by every claim I could not verify. The pattern repeated so cleanly I stopped calling it cleanup: output nobody reads is not output, it is volume.

On Wednesday I audited every scheduled task running for Evalve and rebuilt the set. Six are live now, and the design rule came out of a failure rather than a plan.

Here is the failure. A weekly task researches conference calls for proposals. For every conference it found, it created four Apple Reminders: draft ready, draft due, hard deadline, notification check. Reasonable on paper. In practice they piled up and got deleted unread. Four reminders per conference, roughly twenty drafts produced since summer, one actual submission. The task was working. The output was not.

So I rewrote it to produce exactly one reminder. It is called "Proposal block," it is due Wednesday at 7:30 PM, and it carries five lines: submit tonight, new this week, coming up, check for a decision, and send one podcast pitch. Skip weeks add a line saying no new conference this week instead of going silent. One reminder, one block of time already on the calendar, one place to look.

Then I applied the same rule to the other five:

Midday email triage runs weekdays at 11:21 AM and touches red items only. It drafts, it never sends. Afternoon wrap-up at 3:54 PM drafts the red and yellow replies and names tomorrow's top three. A daily small move at 4:45 PM pushes one five-minute action to my phone and keeps a streak count, because the days I want to do nothing are the days a small move matters. A drive interview queue at 6:08 AM writes three questions I answer out loud on the commute, and a harvest at 8:37 PM turns that recording into a quote bank, a social draft, and a newsletter section. The sixth has no schedule at all: when a person emails Evalve, n8n fires a task that drafts a reply and pushes it to my phone within minutes.

Two constraints shaped all of it. Anything going to a prospect, a client, or a new contact stays a draft until I read it, which takes me a couple of minutes and keeps my name on what goes out. And the file connector I use can create files but cannot edit them, so every task writes a fresh dated file instead of appending to a log. I designed around that rather than pretending it was not there.

One gap I am naming instead of burying: the 2:30 AM triage runs in the cloud and almost certainly has not been reading the Evalve inbox, because the mail connector it needs lives on the Mac. It has been reporting cleanly the whole time. That is the fourth time this year a check told me a verdict instead of telling me what it saw.

Petaluma City Schools Draft Tiered AI Plan for Students by Amelia Richardson, Petaluma Argus-Courier

A California district put its student AI rules on a grade-band ladder: none in K through 2, a basic explanation of how AI works in grades 3 through 5 with no student-facing tools, closed school-only environments in 6 through 8, and teacher-discretion use in high school. Most district AI policy I read is one paragraph trying to cover thirteen grade levels. This is the first draft I have seen that admits a second grader and a junior are not the same decision. Borrow the structure, not the specifics.

Understanding the Evidence Base on AI in K-12 Education by the SCALE Initiative, Stanford Graduate School of Education

They reviewed more than 800 papers and found 20 they would call high-quality causal studies. Zero of those studied student AI use in a US K-12 classroom. Their other finding is the one I keep thinking about: performance often improves while students have the tool and the results get mixed once it is taken away. If a vendor tells you the research supports their product, this is the document to read before you answer.

AI in Education: A 2026 Snapshot of Growing Use and the Shift Toward Integration by Michigan Virtual

Teacher-reported classroom use more than doubled from 2024 to 2026, and trust fell over the same stretch, from 58.4 to 46.5 on their scale. Meanwhile 72.2 percent of administrators say AI is in their district plan, up from 30.6 percent in 2024. Use is up, plans are up, trust is down. That is a training and communication problem wearing a technology costume.

Twenty minutes, and it works on your AI tools, your dashboards, and your meetings.

Step 1. By hand, from memory, list every recurring output your setup produces on a schedule. Reports, digests, reminders, dashboards, weekly summaries. From memory on purpose: the ones you cannot remember are already telling you something.

Step 2. For each one, answer two questions with facts rather than feelings. When did I last open this? What did I do differently because of it?

Step 3. Sort into three piles. Read and acted on. Read and ignored. Never opened.

Step 4. For everything in pile three, you get two moves and improving it is not one of them. Delete it, or fold it into something that already gets read. An output nobody opens does not get better by getting better.

Step 5. For pile two, cut it down to the single line that would have changed a decision. If there is no such line, it belongs in pile three.

Step 6. Write one sentence naming what you cut, and date it. That sentence is your baseline the next time someone proposes a new weekly report.

If the output is AI-produced, this prompt does step 2 for you:

The measure is whether behavior changed, not whether the job ran. That is the day-90 checkpoint from my 30/60/90 AI Implementation Plan, pointed at my own tools instead of a district's pilot. Four reminders a week ran perfectly for months and changed nothing. One reminder gets read.

That's it for this one. If something here was useful, share it with someone who'd get something out of it. If it didn't, that's part of the deal too. You can find more of what I'm working on at evalveconsulting.com, or book a call if you want to talk through what you're dealing with.

Talk soon,

Chris