Opus 5 got smarter.
Your workflow didn’t.
If Claude suddenly argues with you, skips the steps you asked for, and blows straight past your CLAUDE.md — you’re not imagining it. The model didn’t get worse. It got sharper, and a sharper model punishes a loose process. skAIld is the process.
- Tuned for Opus 5
- Terminal or the Claude app
- Every skill included
> add rate limiting to the login route
Before implementing, a few concerns with this approach:
1. Rate limiting at the route layer is arguably the wrong
altitude — a proxy would be…
2. Have you considered whether the login route
is even the right chokepoint here?
3. There are tradeoffs worth surfacing first…
⋯ 9 more paragraphs · 0 files changed
> add rate limiting to the login route
/quick-plan
scope: 2 files, 1 test · plan written
/start-coding
✓ rate-limit.test.ts +34
✓ middleware.ts +18
/create-pr
✓ #128 opened · review clean · 2 files changed
Same model. Same prompt. Same hour. The difference is the workflow around it.
Four things everybody started saying at once.
These are real, they’re reproducible, and none of them mean the model got dumber.
“It argues with me now.”
You say “I think blue is the better call.” It returns fourteen reasons blue is wrong. More reasoning means it surfaces every counterpoint it can find — useful in a design review, exhausting on a decision you already made.
“It skips my steps.”
You asked for a five-step process and got the final answer. It’s tuned to keep momentum and skip the ceremony — which is great, right up until the parts it dropped were the parts you needed.
“It ignores my CLAUDE.md.”
Your rules are right there in the file. It read them, then out-reasoned them. Instructions that live only in prose lose to the model’s own plan unless something structurally re-asserts them.
“It burns tokens going in circles.”
Re-litigating settled decisions, re-deriving context you already agreed on, three days of spend on a two-hour change. Nothing bounds the session, so it re-thinks everything, every time.
The actual diagnosis
“How you use Opus — in what setting, with what harness, and how you talk to it — has a lot to do with whether you experience this.”
That’s the whole thing. Opus 5 is the strongest coding model that has ever shipped. What changed is that it stopped tolerating a loose process.
A vague prompt used to produce a vague draft you’d fix in the next turn. Now it produces a confident, well-argued, deeply-reasoned run in a direction you never asked for — because the model finally has the horsepower to get a long way down the wrong road before you look up.
Smart doesn’t mean it knows the rules of your repo. You have to hand it the rules, every session, in a form it can’t out-reason.
A better way to work with a smarter model.
Not prompt tricks. Structure — loaded at session start, enforced through the loop, so the model works your way instead of its own. You get all of it, on whichever surface you use.
It argues instead of acting
karpathy-guidelines
Four behavioral guardrails loaded into every session: think before coding, simplicity first, surgical changes, goal-driven execution. It stops re-opening decisions you already closed.
It forgets your rules by turn three
/start
A session-start ritual that re-asserts your workflow every single time — environment check, your rules, and the menu of what to do next. Structure beats prose.
It runs off before you agreed on anything
/quick-plan·/start-planning
Decide once, write it down, then execute against it. A spec the model can’t drift from is the cheapest token you’ll ever spend.
It skips the steps you asked for
/start-coding
TDD discipline and a real execution loop — one step, verified, then the next. No jumping to the answer and dropping the middle.
It hands you a giant unreviewable diff
/create-pr·/review-pr·/team-review
Isolated worktrees, scoped changes, and an independent review pass before anything lands. Review→fix→re-review runs in one window.
It ships without checking anything
/pre-deploy·/light-deploy·/handoff
Gates before release, a fast safe path when you’re confident, and a clean handoff so the next session doesn’t start from zero.
Put it in your next session.
One subscription, every skill. Checkout connects your GitHub and hands you the install command — the next session starts inside the loop.
- Session-start ritual + workflow menu
- Karpathy guidelines loaded every session
- Plan → build → ship, end to end
- Updated as fast as the models move
Cancel anytime. What’s included