My Claude Code 'ADHD Executive Function Mode' Went Silent for 6 Days

A monitoring layer that only works when you remember to check it is not a monitoring layer. It is a second thing to forget.

I learned this the way I learn most things about my own harness: by accident, a week late, while looking for something else.

TL;DR

My Claude Code harness logs every meaningful decision to a self-ledger, a running file that acts as my agent’s working memory and my own executive-function prosthetic. Between June 2 and June 7 it wrote nothing. Six days, zero entries, no alert. I only noticed on June 8 when a recall query came back thin. The fix was not “remember to check the ledger.” The fix was making silence itself trip an alarm. That is the whole lesson of harness engineering for an ADHD brain: the prosthetic is the easy part, the failure detector is the product.

Last week vs this week

Last week’s report took apart the “Claude is nerfed” thread and landed on one idea: capability now lives in the harness layer, so two people on the same model are no longer running the same system. (Is Claude Nerfed, or Is Your Harness Flat?)

This week my own harness proved the same point from the embarrassing direction. The model was fine. The harness around it had a hole, and the hole was invisible from the inside, exactly the failure mode I warned about. Invisible is indistinguishable from broken.

What a self-ledger is, and why it is an executive-function prosthetic

People keep searching for “claude adhd executive function mode” as if it were a toggle Anthropic shipped. It is not a feature. It is something you build.

Here is the actual mechanic. My harness writes a structured row to a decisions log every time it makes a call worth remembering, a routing choice, a ratified decision, a goal set, a goal cleared. Future sessions read those rows back before acting, so the agent does not re-litigate settled questions or forget what it was doing across a context reset.

For an ADHD brain, that ledger is doing the exact job my working memory does badly: holding the thread of “what did we decide, and why” across interruptions. The agent’s persistence becomes my persistence. When it works, I stop paying the tax of reconstructing my own reasoning every morning.

Which is precisely why it failing silently is so dangerous. A prosthetic you have come to rely on is worse than no prosthetic when it fails without telling you, because you have stopped keeping the backup in your own head.

The receipt: six days of nothing

I did not estimate this. I counted the rows.

Date	Decision rows logged
2026-06-01	6
2026-06-02	0
2026-06-03	0
2026-06-04	0
2026-06-05	0
2026-06-06	0
2026-06-07	0
2026-06-08	18

June 1 logged six decisions. Then nothing for six days. Then June 8 came back with eighteen, as if nothing had happened. No error surfaced during the gap because nothing in the system was watching for absence. The writes simply stopped, and a stopped write produces no output to notice.

The work never stopped, which is the unsettling part. More than 180 sessions ran on my machine that week, all doing real work, none of it reaching the ledger. I was heads-down building my founder-OS, not auditing the one system whose entire job is so I do not have to. That is the ADHD trap stated plainly: the thing I built to cover the gap failed inside the exact gap it was built to cover.

The root cause was mundane, which is the point. My harness copies the ledger into a database, and that sync stops advancing the moment it hits one opaque production error, unless you explicitly tell it to skip and continue. On top of that, the file watcher that triggers a sync was only watching one folder, so decision-only writes never tripped it. Two quiet failures stacked, and neither one raised its hand. Mundane causes are the ones that run for six days, because dramatic failures page you and quiet ones do not.

Why “just check it” was the wrong fix

My first instinct was the ADHD-brain instinct: add it to the routine. Check the ledger every morning. Build the habit.

That fix fails for the same reason the ledger exists. If I could reliably hold a daily check in working memory, I would not need an external memory system in the first place. “Remember to verify your memory aid” is a contradiction. You cannot patch an executive-function gap with more executive function.

The correct fix moves the work off me and onto the harness. The system already knew how to write rows. It just needed to notice when it stopped.

The 4 things harness engineering for an ADHD brain actually requires

This is the framework I pulled out of the incident. It generalizes past my setup to anyone building an agent they intend to lean on.

1. The prosthetic. The thing that does the job your brain does unreliably, the persistent ledger, the memory file, the routing rules. This is the part everyone builds and the part everyone thinks is the whole job. It is the easy part.

2. The liveness check. A separate, dumb signal that the prosthetic is still running. Not “is the output correct,” just “did it produce output at all in the expected window.” Silence has to be an event. A gap of N hours with zero writes should trip something. The heartbeat is cheaper than the thing it watches and matters more.

3. The recovery. When the liveness check fires, you find the gap in hours and backfill it, instead of stumbling on it a week later. I want to be honest about my own case here, because the flattering version is a lie: nothing self-healed. I had no detector yet, so the system could not catch itself. I found the freeze late, by accident, while doing unrelated work, backfilled what I could, and then fixed the cause. The autonomous version is the goal, not what happened. What happened is the argument for building the detector.

4. The retro that feeds the next failure. The incident only pays off if it changes the harness so this class of failure cannot recur the same way. Every arc has to leave the system smarter than it found it. This article is part of that step: the lesson is now written down where the next session will read it.

The prosthetic is the easy work. The other three are the product. Miss any one and you have a memory aid that works until it does not, then quietly stops carrying you.

The AI-visibility tie-in: silent gaps are the whole game

This is a visibility report, so here is the connection, because it is not a stretch.

Everything I do for AI visibility is the same shape as this ledger failure. A site gets cited by AI answer engines through a stack of signals, structured data, clean crawler access, entity authority, that you cannot see working from the front end. When one of them silently breaks, your pages keep loading fine and your citations quietly stop. No error. No alert. Just absence, for as long as nobody is watching for absence.

The discipline is identical: build the signal, then build the thing that screams when the signal goes dark. Most sites have the first and skip the second, which is why most AI-visibility problems are discovered the way I discovered my ledger gap, late, by accident, while looking for something else.

This week’s chudi.dev visibility numbers

Keeping the series honest means showing my own receipts, not just the lesson.

Metric	This week (Jun 9-15)	Last week (Jun 2-8)	90-day
Impressions	32,190	11,205	94,277
Clicks	294	100	571
CTR	0.91%	0.89%	0.61%
Avg position	6.5	7.2	7.8

Impressions nearly tripled week over week and average position improved from 7.2 to 6.5. Good. But the sitewide CTR sitting near 0.61% over 90 days is its own silent gap: people see the pages and do not click. The single worst offender is the query “claude adhd executive function mode,” 155 impressions at 1.29% CTR sitting at position 9.5, which is part of why this post exists and carries that exact title. The demand is real and the click is leaking. That is a content-and-title problem I can fix, and I am fixing it in public.

FAQ

Does Claude have an ADHD executive function mode?

No, it is not a built-in feature. “Executive function mode” is something you engineer on top of Claude Code: a persistent decision ledger plus memory files that hold the thread of your work across sessions, doing the job that ADHD working memory does unreliably. The model supplies the reasoning; the harness supplies the persistence.

What is a self-ledger in an AI harness?

A self-ledger is a structured log where the agent records every meaningful decision, routing choices, ratified decisions, goals set and cleared, so future sessions can read them back instead of forgetting or re-deciding. It functions as shared working memory between you and the agent across context resets.

Why did the ledger fail without an error?

Because nothing was monitoring for absence. The writes stopped, and a stopped write produces no output, so there was no error to catch. The fix is a liveness check: a separate signal that treats a gap in expected writes as an event worth alarming on.

How do you monitor an AI agent’s memory layer?

Add a heartbeat: a cheap, separate check that confirms the memory system produced output within an expected window and fires when it does not. Pair it with a recovery step you own: find and backfill the gap, then fix the cause so it cannot recur the same way. Automating that recovery is the goal, but the heartbeat that makes the gap visible at all comes first.

How is this relevant to AI visibility?

AI-visibility signals, structured data, crawler access, entity authority, fail the same silent way: pages keep loading while citations quietly stop, with no error. The same discipline applies: build the signal, then build the monitor that screams when it goes dark.

What to do next

1. Find the one external system you have started to depend on, a memory file, a sync, a cron, anything you would not notice failing for a week. Add a liveness check to it today. Not a correctness check, just “did it run.”
2. Make silence an event. Whatever your monitoring is, confirm it fires on absence, not only on errors. Most monitoring only catches the loud failures.
3. If you are building an AI harness you intend to lean on, write the failure detector before you add the next feature. The prosthetic is the easy third.
4. Want the same lens turned on your website’s AI visibility, the signals that fail silently and stop your citations? That is exactly what I audit. Audit your site or start with the Chiron Seed if you want the harness-discipline starter.

---

This is week 2 of the Weekly AI Visibility Report. Last week: Is Claude Nerfed, or Is Your Harness Flat? Next week: whether the “claude adhd executive function mode” retitle moved the click.