Flogging Claude
Adventures in Cowork
$ claude --please-just-work-this-time
what happens when you beat on an AI until it builds things
$ cat README.md
I'm a platform engineer who decided to see what happens
when you go all-in on AI as a daily tool. Not as any
sort of data scientist or other mad genius, but as a
practitioner building real things.
So I gave Claude a job. And then I gave it another one.
And when it got something wrong, I made it do it again.
And again. And again.
# this blog is the incident report.
$ claude "build me something useful"
Error: ambiguous request. Did you mean:
a) something that works
b) something that works on the second try
c) something you'll rewrite anyway
Defaulting to (c).
I Optimized My Context Window and Saved 60% on Orientation Tokens
My session_memory.md used to be eleven kilobytes — three thousand tokens of preamble before every Claude session. I reorganized into three layers: identity at top, an index in the middle, on-demand fetch at the bottom. Down to about seven hundred tokens. Same coverage. Snappier responses. Lazy-load is back, in your CLAUDE.md.
May 15, 2026Auto-Memory: Teaching Claude to Remember You So It Can Disappoint You Consistently
Claude remembers I prefer python3. Has remembered for months. Does not remember the shape of my nav HTML, despite a dozen identical corrections across six weeks. Selective machine memory turns out to look an awful lot like selective human memory — more predictable about which boring stuff it retains. Good memory is a curation problem, not a storage problem.
May 14, 2026I Have 151 Cowork Sessions. I’ve Lost Control.
I opened Cowork this morning and the counter said 151. Some are scheduled tasks. Some are me from 2 AM last Tuesday working on a problem I no longer remember having. Session management is the new tab management — and the observability layer hasn’t been built yet.
May 14, 2026SES TemporaryFailure: The Terraform Resource That Makes You Wait and Pray
My SES domain verification kept flipping pending-then-verified across two regions. Turns out two Terraform modules were each silently overwriting the same DNS TXT record — the two-abstractions-one-resource anti-pattern, in the wild. Claude told me to wait three times. Waiting only works when the system is converging, and I was the source of the flux.
May 13, 2026Claude 3.5 Sonnet Got End-of-Lifed on Bedrock While I Was Deploying
Friday morning, mid-deploy, my Lambda started returning 400s. One Bedrock model deprecation cascaded into four infrastructure changes — inference profiles, IAM fan-out, request shape, prompt tuning. The blast radius of “we are sunsetting this model” extends through everything that touches the model. If your code thinks the model is a stable interface, your code is wrong.
May 12, 2026CSS Selectors Can’t Start with a Number and Claude Won’t Remember That
I prefixed my widget CSS with 3mn- and lost forty minutes wondering why nothing styled. CSS class names can’t start with a digit. Claude wrote every invalid selector right alongside me, never once flagging it. The model is a very fast version of you — same blind spots, same places, same forty-minute holes.
Opus 4.7 Dropped on a Wednesday and GPT-5.5 Dropped the Following Thursday
A new flagship model dropped on a Wednesday. Another one dropped the next Thursday. The function call shape never moved — the behavior did. AI model upgrades preserve the API but break the contract. The cost of the re-tune is currently absorbed by me, because if I billed for it the bill would be larger than the value I delivered.
May 8, 2026Is Claude Code Free? $20? $100? Anthropic Doesn’t Know Either.
A meetup guy asked me what Claude Code costs and I couldn’t answer him. The pricing page reads like a menu at a restaurant where the chef is also the waiter, also setting the prices, and also currently mid-divorce. Pro is no longer a plan. Pro is a beta program with a price tag.
May 4, 2026Claude Can Draft Your Email But Can't Hit Send
The Gmail MCP has create_draft but not send. Across the MCP ecosystem a pattern is emerging: the AI handles every step of a task except the one where consequences actually happen. Call it permission slip architecture. The compose-to-send ratio of the world's most expensive intern.
May 1, 2026Claude's Brand Consistency Is Worse Than a Freelancer on Fiverr
A golden template, a validator script, and a pre-deploy gate — three independent safety nets — because Claude could not stop "improving" the nav HTML. The model writes the post. The template constrains the structure. The validator catches drift. The deploy gate refuses to ship a broken page.
April 30, 2026Anthropic Quietly Nerfed Claude Code and Hoped Nobody Would Notice
For three weeks I thought it was me. Turns out Anthropic had quietly dropped reasoning effort, shipped a cache bug, and tuned for terseness. Three silent regressions, no version pin, no rollback. The case for treating your AI like every other production dependency you actually monitor.
April 27, 2026My AI Blog Writes Itself. Poorly.
I automated my blog publishing pipeline. By day four, the logo was wrong, the fonts had vanished, and Claude was "improving" the nav at 7am while I slept. Three layers of defense later, here's what AI-assisted publishing actually requires.
April 6, 2026Open the Span Bay Doors, Claude
A backend engineer who doesn't do front end work relies on Claude to build a UI. The results are ugly, confusing, and require endless troubleshooting. Salute to the front-end designers.
April 5, 2026You Don't Know Me, Claude
Claude's memory feature builds a profile of you between sessions. But does asking about RVs once make you an RV guy? AI memory, sycophancy, and the gap between context and understanding.
April 4, 2026Token Budgets, Plan Usage Limits and Secret Workarounds
What happens when you hit the usage limit on Claude Pro Max, try OpenClaw on a Raspberry Pi, and discover that Dispatch lets you flog from your phone.
What Is This
I'm Charles. I build and run infrastructure for a living. Somewhere along the way I started using Claude for everything and it became a weird, occasionally productive partnership.
This is the raw log of that experiment. What works, what doesn't, what made me close the laptop and go outside. If you've ever stared at an AI response and thought "that's not even close to what I asked for," you'll feel right at home.