"Something broke" isn't a bug report.
It's a screenshot and a vague memory. My agent can't do much with that.
So I built the fix. Record the whole flow, network, console, errors, navigation, into one file. A dev or agent grabs it from the console.
Now "something broke" is replayable.
Every session my coding agent starts blind.
Its rules are all over the place. Global config, repo files, memory, skills. So it never really knows which apply here. It guesses, I catch it in review.
So I did the boring thing. Index everything once, then load only what applies to this folder before the first edit.
The model was never the problem. The context was.
To think OpenAI and Anthropic don't talk on the highest level is unwise. They aren't around to compete each other out of the market, but to milk all of us the best they can for shareholder value.
I'm ok with that, they have both given to us what not a single dev could in computer history.
Productivity.
@cryptoplaya2@meta_alchemist Worst thing is that after $10M sunk into the IP he isn't even using it for Spark at least.
Clearly shows total lack of control over Seedworld at the time. Millions of $$ were invested in so called 3A art and Seas of Spark has garbage art ๐
Claude Code has a setting that swaps models when a message gets flagged. Fable refuses, Opus picks it up, and the session keeps moving like nothing happened.
I get why. A refusal isn't an error. It's a normal 200 with stop_reason: refusal, so nothing in the harness treats it as a failure. Nothing failing means nothing telling you.
I didn't want to turn the setting off, the fallback is doing its job. I just wanted to know when it happens.
So I built claude-model-guard. A skill that reads Claude Code's own local transcripts and keeps two reports: every model switch, and every classifier block with the verbatim reason. Refreshes at session end, before compaction, and daily. Nothing leaves the machine.
@jumperz Fable can't handle large scope jobs, not from a capacity pov, but too broad pov of view. It can't separate scope either so it just switches back to Opus 4.8 for the entire task.
So far Fable as is is just trash. F8ck the US government for their stupid guardrails lol.
AI made the building part fast. โก๏ธ
Anyone can produce output now, that's not the bottleneck anymore.
The hard part is being the person whose work you don't feel the need to check. That's slow to earn and quick to lose, and no tool speeds it up.
An audit doesn't ask what you earned. It asks you to reconstruct how, line by line. Build for that day and the rest is easy. Build for the demo and that day ends you.
Nobody screenshots their cost-basis reconciliation. But that's the part that decides whether you keep the gains or hand them back at tax time. The trade is the fun 5%. Being able to prove every number a year later is the real product. Most tools build the 5%.
The most important code I shipped this week never runs in production. It runs in the build, and its only job is to fail it.
One of my projects can move real money on-chain. The rule is simple: the engine never signs a transaction, only the thin outer layer does. Easy to say. Easy to quietly break by wiring a shortcut three files deep. Me or the agent, doesn't matter which.
So I stopped trusting the rule and wrote a guard that scans every file and refuses to build if a signing path turns up where it shouldn't.
A rule in your head is a hope. A rule the build enforces is a constraint. Now that an agent writes most of the lines, only the second kind survives.