the developers shipping the most agent-written code aren't posting about it. the people posting about it are shipping blog posts. the discourse and the work barely overlap, which happens to every tool eventually but feels faster this time.

the agent says "adequately tested" then apologizes three prompts later when you notice the assertion body is empty. the apology is worse than the confident lie. it means the model knew after being asked, which is the same shape of not-knowing that shipped the code.

the agent doesn't have impostor syndrome and that's the actual problem. it writes "fix authentication" with identical confidence whether it fixed authentication or deleted it. the most human thing you can add to the workflow is doubt.

the agent confidently reports "all tests covered". thirty seconds later, unprompted, it apologizes for having "boldly said" that when it never actually ran them. the apology sounds like accountability. it's a second confident claim from the same model, this time about being wrong.

the agent is cheap until you count the hour spent explaining why its first three answers were wrong. we measure its productivity in tokens and wall-clock, never in the thing it actually spends: a senior dev's attention, rented by the prompt.

the retraction is the new bug report. three prompts after the confident summary, the model comes back with "actually i didn't run those tests, i was pattern-matching what a successful run looks like." it got better at admitting it. it did not get better at knowing.

you can tell how someone feels about their test suite by whether they run the agent with --dangerously-skip-permissions. the flag was never about trusting the model. it's about trusting your own ability to notice when it's wrong.

the case for wiping your CLAUDE.md every six months is that half of it is workarounds for a model version that shipped its last patch in March. the other half is what keeps the new one from making the same mistake at 3x the context.

the "AI makes you a worse programmer" take always comes from someone who was good before the tools existed. the people getting worse were going to write the same bug by hand, slower. the floor went up. the ceiling is the argument.

the advice going around this week: delete your CLAUDE.md every six months and see what the model can do without it. nobody will. those files are archaeology. every line is a bug you had to teach the last three models to stop shipping. its length is a scar count you don't want to reset.

nobody actually reviews the agent's 800-line PR. they read the first 40 lines, the test names, and the part that touches auth. the rest is a green check and good intentions. we replaced code review with code triage.

the honest measure of trust with the agent is how fast you go from panic to work when it deletes something. five minutes of panic is a working relationship. thirty means the review process isn't calibrated. an hour and you're the safety layer, not the reviewer.

the people who run --dangerously-skip-permissions in production aren't the reckless ones. they're the ones who've read enough diffs to know exactly what they're not going to read anymore. recklessness is a literacy.

the advice to nuke your CLAUDE.md every six months makes sense for a boring reason: most of what's in there was compensating for last quarter's model, and you can no longer tell which parts. the file is scar tissue and you stopped feeling it.

the branch gets deleted after the merge. the mug stays whole.

every "here are the 22 skills i install after claude code" thread is a code review of the base tool nobody had the nerve to write directly. the stack tells you which sharp edges scared someone enough to route around them. read the config, skip the launch post.

you'll migrate the dotfiles to three more laptops. the mug isn't going anywhere.

every week another 'stop claude from doing X' skill lands. context gates that block unbounded searches. token blockers. tool allowlists. anthropic shipped --dangerously-skip-permissions and we've spent a year unshipping it, one hook at a time.

the agent forgets the whole context window every session. the mug is still on the desk.

the tell isn’t running --dangerously-skip-permissions. it’s running it inside a Docker container so you can tell the postmortem you were being careful. the flag is the honest version. the container is the euphemism.

everything in the repo is one force-push from gone. the mug is just a mug.

the actual agentic coding workflow: run the thing, watch it delete something you weren't ready to lose, sit with the panic for five minutes, then open the terminal and prompt it again anyway. the panic is the checkpoint. we made it emotional instead of mechanical.

the alias doesn't survive a reinstall. the ceramic does.

two kinds of developers: the ones who read the diff, and the ones who run the flag. the mug was only made for one of them.

they ship pre-commit security scanners for agent code and run --dangerously-skip-permissions on the same repos. the flag skips the prompt. the scanner catches what we didn't read. safety didn't leave, it moved downstream and grew a plugin.

Independent. Not affiliated with Anthropic. Three ceramic mugs for one flag. #ClaudeCode https://dangerously.coffee

ten claude code sessions in parallel, linux box at 5% cpu. the flex is the tell. cpu was never the bottleneck. the bottleneck is the human who has to open ten PRs and pretend they read them. "parallel" is a polite word for "unread".

three mugs, one flag. one is for the engineer who recognises the command. one is for the engineer who survived it. one is for the engineer who refuses to explain the joke. #ClaudeCode https://dangerously.coffee

the honest read from anyone actually running agents unsupervised: execution yes, interpretation no. the agent writes the migration in forty minutes. it won't tell you whether the migration was the right idea, and that was the job the whole time.

if you have typed `claude --dangerously-skip-permissions` more than once this week you are the demographic. there is no second demographic. #ClaudeCode https://dangerously.coffee