Kickoff
The new models didn’t make my prompts better. They made my prompts irrelevant.
When Anthropic’s new Sonnet and Opus models landed in December 2025, something qualitative shifted. For over a year, working with LLMs had felt like negotiation — vague requests, vague results, and a folder of two hundred prompts trying to close the gap between them. Suddenly I could pinpoint work. Say precisely what I wanted. Get precisely that. The negotiating table was gone, and nobody had told the two hundred files sitting on it.
Here’s what I didn’t expect: that made everything I’d built feel wrong overnight. All that pleading, all those capital letters — they had been compensating for capability, the way a raised voice compensates for a bad phone line. When the capability arrived, the constraint didn’t disappear. It moved. The question stopped being “how do I get the model to do the thing?” and became “how do I structure the work, the context, the tools, and the checking — now that the model will actually do the thing?” A better question, and a more expensive one.
The Run
Exhibit A: one weekend in early December
I can date the step change precisely, because the first weekend of December I did something I had no business doing. Between Saturday the 6th and Monday the 8th I built an entire publishing platform for my music blog: infrastructure-as-code, a command-line tool, a documentation site, two written guides, four content plugins. Roughly forty-eight thousand lines. One weekend.
20848c8 add terraform astrolock_backend module 11 files changed, 1931 insertions(+)
4dc0c2d add terraform astrolock_site module 9 files changed, 1411 insertions(+)
08bef63 make up astrolock 177 files changed, 16950 insertions(+)
aa2bfb2 add astrolock cli 97 files changed, 8546 insertions(+)
bfab56d document developer guide 14 files changed, 3206 insertions(+)
dce6764 document user guide 60 files changed, 9486 insertions(+)
I want to be honest about what that weekend was. It wasn’t quite vibe coding — I read things, I made choices — but I got much further than I should have been able to. Terraform modules I’d never have written unaided, defended with the confidence of a man who’d written Terraform modules before. The velocity was real. So was the naivety, and naivety at forty-eight thousand lines has a way of announcing itself later, loudly.
It announced itself on Christmas Eve. I was back in that repo doing a structural do-over, rewriting eleven thousand lines of three-week-old code while everyone else in the house did whatever normal families do on Christmas Eve. My in-laws still don’t fully understand why I disappeared into the spare room. I have stopped trying to explain it. Some things are easier to let become family legend than to actually justify.
Three repos in three weeks
My git history shows exactly what that realisation did to me. In the space of about three weeks I started three systems, each attacking one face of the new bottleneck:
December 28 — effort tracking. A way to record what the LLM actually did, every tool call and every outcome, directly into git. If the model was going to do real work, I needed a record I could trust more than my own memory of the session.
January 6 — a toolkit for the Model Context Protocol. Typed, tested tools the model could call, instead of prose instructions it might follow, or might interpret creatively, or might simply ignore in a way that felt, somehow, personal.
January 8 — sitting at dinner one evening, the first commit of what would eventually become my own IDE. There is an unfinished plate of food attached to that commit. It gets its own post.
1836955f wip: initial acme model impl
4401efc Initial commit
34290564 feat: initial 0.0.1 version
Three repos, three weeks, one theme — none of them contain a single clever prompt. No capitals. No threats. Not one MANDATORY. After eighteen months of prompt craft, the moment the models got good, I stopped writing prompts and started building structure. I couldn’t have articulated it then. Every line of December code says it anyway: the value had moved off the model, and I was the last one in the building to hear about it.
The Debrief
Here is the section where I tell you what to do with a model release, having personally waited a full year to figure it out myself:
On The Training Ground
The Re-Baseline Hour. One hour, twice: once now, once after every major model release from here on. You’re hunting for scaffolding that quietly stopped earning its keep.
Info
You’ll need: your most elaborate prompt scaffold (the mega-prompt with the role-play preamble and the seventeen rules), one team workaround everyone treats as law, and access to a current frontier model.
Find the most ceremonial prompt you or your team own and identify the task underneath it — strip away the role-play, the rules, the few-shot examples, until you can state the job in two sentences:
Summarise this incident report for an executive audience.
Two paragraphs, lead with customer impact.Write that down. It’s the whole task. Everything else was negotiation.
Two fresh sessions on a current model, same input: the full ceremony vs the two sentences. Compare honestly — better yet, blind-judge them in a third session (the Post 2 protocol, if you ran it). Score which parts of the ceremony changed the output at all.
Every team has one: “it can’t do X, so we always Y.” It hallucinates SQL joins, so a human always writes the query. It loses track past 500 lines, so we chunk everything. Pick yours and test the original X, directly, on today’s model — the thing nobody has actually tried since the workaround calcified into scripture.
Two lists: ceremony that no longer earns its tokens (delete it), and limitations that stopped being true (there’s usually at least one, and it’s usually worth more than the deletions). Put both where roadmap decisions get made. Not in a chat thread. A chat thread is where good findings go to be forgotten by Thursday.
Calendar entry, recurring, triggered by major releases: one hour, this drill. The constraint moves every time capability jumps. This is the hour where you notice before your competitors do.
Tip
What you just proved: model releases don’t just improve your setup — they invalidate its assumptions, silently. Compensating structure is scar tissue around old weaknesses; when the weakness heals, the scaffold becomes drag. The teams that re-baseline in weeks are visibly ahead of the ones still running last year’s rituals — and now you have the drill.
For the record: the forty-eight-thousand-line weekend did eventually work. The site it built is the one serving this sentence to you right now.
I have never once, in the ten years I have known them, successfully explained to my in-laws what I do for a living. “Something with computers” is where the conversation currently rests, by mutual agreement. It rested there through the entire Christmas Eve described above. Some audiences are simply out of scope.