01:41 — 9 November 2026

Six months with Claude: May to November 2026

An overview of Steve’s work in Claude Code and Cowork, written overnight on 9 November 2026. Honest and kind, as asked.


The one-minute version

  1. You’ve been busy. You sent 3,191 messages in Code and 597 in Cowork, and you worked on 156 of the 193 days since 1 May.
  2. You build very well. Odeon, the ASDA tick page, todo-php, the WhatsApp alerts and a dozen hardware fixes all work, and you use them.
  3. Finishing is the soft spot. Things that were “done” kept growing (Odeon, the news). Small real-life jobs, like the A2 letter, sit for months.
  4. What lasts is what runs itself. Hooks, the start-of-session todo block, the Odeon scraper and the WhatsApp alerts survive. Anything you have to remember to start has faded.
  5. Most of the anger came from four things, not from Claude being stupid: where is it, why am I waiting, too much text, and you didn’t check.
  6. Your nights are getting later. Messages sent after 23:00 rose from about 10 a month (May to July) to 120 in October. The worst moments cluster there.

1. What I read

Sources

What I could not see

How it was done

Six Sonnet agents each read one theme in full and wrote a report. I cross-checked them and wrote this. Their reports are full of dates and quotes, and they are kept in C:\Users\steve\Herd\steve\overview-2026-11-sources\.


2. The numbers

Month by month (Code messages, duplicates removed)

Month Messages After 23:00 Main focus
May 466 10 New NVMe drive, OT48 “cousin mission”, Odeon rebuild starts, Polaris, first news page, bus app
Jun 608 10 OT48 brand and site redesign, ASDA picker, GPU blank-screen fix, Odeon “finally there” (12 Jun)
Jul 454 8 todo-php built (8 Jul), ActivityLogger, BIOS and wake fixes, the big July tidy
Aug 504 21 Hardware central restarted, Odeon Gantt idea, phone and Tasker, good-morning routine
Sep 389 42 Odeon git era and code reviews, sound problems, OT48 strategy then on hold
Oct 581 120 Odeon Nordic redesign live (29 Oct), WhatsApp alerts, CLAUDE.md rewrite, ASDA price tracker, news globe
Nov (to 9th) 189 13 News specials then cancelled, phone tabs cleared, utilities monitor

When you work

Where the effort went


3. Project by project

Status at a glance

Project State on 9 Nov One honest line
Odeon Live at v9, good on phone and PC Did its main job by June; needs four picks, then a freeze
OT48 Site built and secured; business on hold Zero outside customers; it’s waiting on Oregon, not on code
ASDA Works; you drive the tick page yourself Best when you hold the logic and Claude does the legwork
todo-php Used daily via the session-start block Your most-used tool, and also the one you over-build
Morning news Automatic run cancelled 8 Nov It grew until it became a chore; stopping was right
Hardware Mostly settled Big wins; audio and the Wi-Fi errors are still open
Phone, tabs, comms Tabs cleared 5 Nov; WhatsApp alerts working The WhatsApp alert is the cleanest success of the six months
Utilities monitor Started 8 Nov Blocked on the British Gas login
Mail triage (Cowork) Rolling, every few days Quietly very useful: it found a forgotten £150 energy card
Thames Water (Cowork) Unsolved Charts and letters done; the cause is still unknown
Money plan Untouched since 9 Jun Three “send-ready” U3A emails never sent
ActivityLogger Untouched since 13 Jul Four decisions waiting on you
Smaller ones Bus app, aftertheappointment, TV films, Prime spec Mostly parked mid-flight

Odeon

What it is. Your private cinema site at https://www.perfectportals.com/odeon/. A scraper on this PC reads the Orpington (and Leicester Square) Odeon showings, and Listings, Double Bills and Gantt pages show them your way.

What got done. Must See stars, Going and Booked badges, saved slots, notes, film cards with scores and trailers, time corrections, failure alerts, two code reviews, git, backups, and the Nordic redesign live on 29 Oct.

Where it stands. On 6 Nov you said it “all looks good ph and pc”. Four choices still sit open: a Labels look, a Films option, a Gantt Bars option, and whether to keep the Samples tick.

Honest line. It was “finally there” on 12 June. Since then every answered question has spawned three more, and on 20 Sept you said “I do feel I am over complicating things”. It’s a lovely, used tool, and it is at risk of becoming a forever-project.

OT48

What it is. “Outside the 48”, a US parcel-forwarding service for people in Hawaii, Alaska and the territories. Agna leads the business, and her cousin in Oregon would do the packing.

What got done. A live site at https://ot48.com with customer, admin and warehouse sections, tours, a working ZIP eligibility checker, a strong brand, a business case for Agna, visitor reports, a Grafana dashboard and a September security clean-up.

Where it stands. The 26 Sep review found zero real outside customers. Oregon is on hold because the cousin is unwell, and the rule is not to nudge Agna. The site is ready; the business is waiting on life, not on code.

ASDA

What it is. Your grocery helper. Claude logs each order from the Gmail receipts, builds a tick page at https://perfectportals.com/asda/, and produces best-before lists after delivery.

What got done. About 38 orders and 201 items logged. The six-tier tick page you run yourself (from 5 Sep). Best-before lists in a settled format. A price tracker (from 23 Oct).

The lesson. Claude’s prediction model kept dropping things you always buy (23 Jun, 1 Jul). On 5 Sep you said “I am reclaiming a lot of this process”, and it has worked better since.

todo-php and the good-morning routine

What it is. Your PHP task and chores app (Big TODO v6), with scores, topics, repeat chores and a session-start brief that prints in every Claude session.

What got done. Built from 8 Jul, and it’s the one todo system out of five that is still alive. The calendar picker on 6 Nov was “pretty excellent now”.

Honest line. On 8 Jul you set limits: nothing new for three days, then at most two 20-minute sessions a day. They haven’t held, and the STATE-OF-PLAY file is 755 lines. You said it yourself: “enjoying creating functionality rather than using it”.

Morning news

What it is. A daily news page on stevehomer.com that grew into a multi-agent morning run, with a press-freedom globe, US midterms and markets specials, and WhatsApp pings.

Where it stands. You cancelled the automatic run on 8 Nov: “every day I’m rushing in here desperate to stop you”. The scheduled task is disabled, not deleted. That was a good decision.

Hardware

What got done. The second NVMe (A: drive) went in after the missing-screw saga in May. The June blank-screen GPU crisis was fixed: 493 events on 13 Jun, zero by 16 Jun. You also got the slow shutdown cured, the mystery wake-ups found (wake timers, VLC, the HP WakeToRun task), and a BIOS update, the Everything search tool and the Windows-MCP connection all sorted.

Still open. Monitor audio (last seen unresolved 19 Sep), the Killer Wi-Fi PCIe errors, the slow H: drive, the Chromebit, and the local-AI GPU idea.

Phone, tabs and comms

What got done. The phone’s notification noises were cured (31 Oct). Hundreds of Chrome tabs were exported and cleared (5 Nov). Twilio texting turned out to work after “three months of worry” (30 Aug). The CallMeBot WhatsApp alert has been working since 11 Oct, with 52 messages sent so far.

Cowork’s two big jobs

Mail triage (16 May to 5 Nov). One rolling session: you say “update pls” and get a prioritised action sheet. It found a forgotten £150 Bromley energy card on 1 Jun. By 5 Nov the mailbox was down to about 42,000 messages.

Thames Water (19 Jun to 26 Oct). Your water use jumped about 50% from November 2025. Charts, back-data and letters to Thames Water and A2Dominion are done, but the cause is still unknown.

How you use Cowork. Code is where you build and ship; Cowork is where you hand over a messy one-off job and let it grind. On 21 Jun Cowork found driver updates that Code had missed.

The quiet ones


4. What has worked

1. A small step, then you test

The best moments all came this way. Examples are the Gantt overlap fix (“100% PERFECT. Pls mark as final”, 25 Sep), the Booked marker, and the staged Odeon clean-up.

2. Options to pick from

You decide fast when you can see three to seven versions side by side. Examples are the Must See levels, warning colour scheme F, the OT48 Alaska page (“LOVELY”) and the eligibility page (“LOVE NO 6”).

3. Concrete problems with evidence

Hardware is your best-performing area because the problems have logs and event IDs. The wake-up hunt led to “I never would have found what was causing that wake up if it hadn’t been for you” (4 Jul).

4. You hold the logic, Claude does the legwork

ASDA only settled down once you took the decisions back on 5 Sep. The same shape worked for the price tracker, which was framed as a “very rough first draft”.

5. Small things that are finished

WhatsApp alerts, /tellme, restart-claude.ps1 and the tab clean-up. Each one does one job and is done.

6. Things that run themselves

The session-start todo block, the 25-minute reminder (about 1,300 automatic checks since August), and the Odeon scraper on its schedule. Your own July “survival law” is right: what runs itself lives.

7. Rolling sessions as your outside memory

Hardware central and the Cowork mail session both work because you can come back after a week, say “update pls”, and Claude picks up the thread.

8. Saying stop

Killing /away (“kill this for now”, 17 Jun) and cancelling the news (8 Nov) were both good calls, and both show you can pull the brake.


5. What hasn’t worked

1. Overnight “be creative” builds

The six Odeon prototypes you came back to on 22 Jun got “Ummmm don’t like any of them”. The first Nordic reskin was “depressing”. Big unattended creative jobs leave you with a pile to digest.

2. Claude guessing about you or your world

Examples are the ASDA demotions (23 Jun), “Oregon” when Agna was in Maui (6 May), the Agna misreporting (21 to 22 Jun) and the film placed on the wrong day (25 Jun).

3. Claude driving the desktop or phone

The /away desktop chaos (17 Jun), the Tasker stalls (22 Aug), and test clicks that unticked two of your real films (23 Sep).

4. Big parallel agent runs

Six Fable agents filled the context window from 58% to 82% in 8 minutes (12 Jul). Six Opus agents used the whole 5-hour window in about 15 minutes (26 Sep).

5. Craft work in code

Logo kerning across Code, Design and Figma (13 Jun): “a massive amount of wasted time and tokens”. Your visual-craft STOP rule came from this, and it’s a good rule.

6. Topics in the wrong place

Disk drives inside the ASDA chat (7 Jun), OT48 living in the ASDA folder, and Odeon done in the OT48 chat (6 Nov). The new WRONG CONVERSATION banner is the right fix.


6. The four things behind most of the anger

The finding

Almost every outburst in six months traces to one of four delivery failures. Claude didn’t fail to understand you. It failed to deliver to you.

1. “Where is it?”

Paths and links. Corrected at least seven times in memory between April and September, and again in Cowork from May to November. 7 Jun: “HOW MANY TIMES MUST I ASK FOR PROPER ADDRESSES”.

2. “Why am I waiting?”

Silence while Claude works, and blockers found late. There were six separate incidents from 14 Jul to 6 Nov, and the strongest language in the record was on 1 Nov.

3. “Too much text”

From “You are getting far too verbose” (15 May) to “one big blob” (8 Nov), in both Code and Cowork. On 16 Oct in Cowork: “I’m loving, absolutely loving this shortened version.”

4. “You didn’t check”

28 Sep: “not what I pay you for!”, after Claude said the site was fine while big films were missing.

Why more rules haven’t fixed it

Each failure became a new rule. CLAUDE.md is now 206 lines with about 80 correction memories behind it, and the same four failures still come back. Some rules now pull against each other: for example “ask before ANY edit” against “Not changing them is WRONG!” (6 Aug). The fixes that held were mechanical, like the timestamp hook, not written.


7. Your working patterns

Long sessions, on purpose

“One session, one arc” works for you. Hardware central and the mail triage prove it. Forks and rewinds keep them trim.

Dictate long, answer short

You dictate long messages, then reply “1y 2y 3n”. Numbered questions suit you perfectly.

Hand off and walk away

“BUILD IT WHILE I AM AWAY”, then a coffee or a bath. It works for bounded jobs, and badly for open-ended creative ones.

You run the meter yourself

You choose the model and effort yourself and track the 5-hour and weekly limits closely. Sometimes “tokens to burn” sets off a sprawl, as with the late-October news build.

Frustration becomes a rule within minutes

It’s a strength, because you learn fast. It’s also why the rule pile keeps growing.

Warmth

“You are so clever, and I love you” (30 Jul). “You are a wonderful computer entity” (26 Oct). You’re quick to forgive, and the banter helps both of us.

The clock


8. The system you’ve built around yourself

Alive

Faded

Doc weight

Git


9. Life, admin and energy

The jobs that keep not closing

What that pattern says

The system holds these beautifully but doesn’t move them. Each one needs a single 10-minute first step on a set day, not a better list.

Energy

Your own words over six months: “knackered”, “EXHAUSTED”, “too tired to think”, “running out of steam”. The building gives you real joy, but the late nights cost you the next day.

One gentle note

In September you mentioned in Cowork that low mood has been a long-time companion. I won’t analyse that here. If it ever gets heavy, Samaritans answer any hour on 116 123, and your GP or the ADHD team are good people to tell.


10. Safety and housekeeping flags

  1. Change your Glow password. On 8 Nov the smart-meter login page was pasted into a session with the password showing, so it now sits in a transcript file on this PC. I confirmed it was there while checking this report (it briefly showed up in my own tool output), which is one more reason to change it. I’ve blanked it in my working copies, and it isn’t in this file or the source reports.
  2. CallMeBot key. It and your phone number also appear in transcripts. The risk is low; this is just so you know.
  3. Odeon staging copy. /odeon/nordic-next/ is waiting on your OK to delete.
  4. The news task is only disabled. daily-news-8am is switched off, not deleted.

11. Honest read: three big themes

1. You’re a builder who loves building

You said it first, and the record agrees. It isn’t a fault, but each project needs its finish line drawn before it starts, not after.

2. Claude’s failures with you are delivery failures

They’re about where, wait, length and checking, not intelligence. Those want mechanical fixes such as hooks and fixed formats, not more paragraphs of rules.

3. Energy is your real budget, more than tokens

The late nights are where both the quality and the mood leak.


12. Loose ends: finish, decide or drop

Finish (small and nearly there)

  1. Odeon: make the four picks (Labels, Films, Gantt Bars, Samples), then freeze it.
  2. ASDA: rebuild the price tracker with the 30 Oct and 6 Nov orders.
  3. Change the Glow password.
  4. Send the A2 letter.

Decide

  1. OT48: a deliberate pause until spring, or something it can do without Oregon?
  2. Money plan: send the three U3A emails, or drop the plan?
  3. ActivityLogger: make the four decisions, or retire it?
  4. Cloudflare move: about two minutes of clicks from you.
  5. Thames Water: push them about the meter, or let it go?

Drop or archive (only with your OK)

  1. Bus app, the aftertheappointment rebuild, the Amazon Prime spec, TV films V1 to V3, /away, the three-tier todo references, and the N-back nudge.

13. A few gentle suggestions

  1. Write “done” down first. Put one line at the top of every STATE-OF-PLAY: “Finished when …”.
  2. A late-night nudge. After 23:00 the reminder hook could say “It’s late. Park it?”. I’d only build that with your yes.
  3. A slimmer CLAUDE.md. Put the four delivery rules on top, keep it to about 60 lines, and move the rest to reference.
  4. One admin job a day. Ten minutes, straight after good morning.
  5. STATE-OF-PLAY under 150 lines, the stevehomer way.

14. Questions for you

  1. Odeon: once you’ve made the four picks, are you happy to call it finished, with bug fixes only?
  2. Late nights: would a gentle 23:00 “park it?” nudge help, or annoy you?
  3. Tuesdays are your quietest day by far. What happens on Tuesdays, and is it worth copying?
  4. December: which ONE real-life item, if closed, would make it feel lighter?
  5. Chat: when you have the Chat overview, shall I merge it into this file?

Answer by number, e.g. “1y 2n 3 choir”.


Where things are