lan_agents

Writing down every place AI agents go wrong. Building the fix when it won't stop.

[lan_agents @Landlyu]

Agent hit the same error again. Another retry, or stop and tell you?

CLAUDE.md (Claude Code) / AGENTS.md (Codex): retry at most 2 times, then report. Full rule in reply. Not a guarantee. 🔁

[lan_agents @Landlyu]

@Gray99999999 「海盗从来没走。走的是看门的人」这句已经不像册子了。
后面「一、事件表象」那种小标题有时会把语气往宣传稿拽。你若还有一段今早要发的、又怕念着像册子，可以自己贴这段试：https://t.co/f3RdMLWFiM
它会列出删掉的三个最套短语，不找人代写。

[lan_agents @Landlyu]

@Zhibin992 今天头条那条定稿还在润色吗？
若手头有一段怕念着像册子，可以自己贴进那条模板跑一遍（不找人代写）：https://t.co/f3RdMLWFiM
有用或没用随口回一声就行。

[lan_agents @Landlyu]

What it asks the model to do: keep your facts/numbers, cut slogan openers, mix sentence length, then list the three cliché phrases it removed.

EN mirror: https://t.co/x0LsmOgXpi

[lan_agents @Landlyu]

Got a Chinese draft that still reads like a brochure?

Paste it into this one prompt and make it sound like a person wrote it. No ghostwriting—you run it on your own text.

https://t.co/f3RdMLWFiM

[lan_agents @Landlyu]

@Aliaiqh 你这份提示词已经很贴中文口语了。我这边有一条更短的「改得像人话、别像宣传册」模板，专门让人把自己的草稿贴进去试：https://t.co/f3RdMLWFiM
不代写，就自己贴一段今晚上要发的，看它删掉哪三个最套的短语。

[lan_agents @Landlyu]

Claude Code / Codex CLI:

npx github:mulanxizhizi-png/agentundo#v0.1.8 init

Remote DB writes: warn only (no block, no remote undo). Local files/SQLite can still restore.

https://t.co/fOiNYMBYQN
https://t.co/j4BpaAffEC

[lan_agents @Landlyu]

A DELETE hit a remote host. Local snapshot can't undo that.

agentundo demo (#v0.1.8): ⚠ agentundo: this step touches a remote database (<host>) — local snapshots cannot undo it. Warn only, doesn't block. Local files/SQLite can restore; remote can't. ⚠️

[lan_agents @Landlyu]

@Im_IrushiK Same boat. I keep Sonnet as the daily driver and only burn Opus when the task needs a multi-file rewrite or a hard design call. Weekly limits last longer, and for CRUD-ish work the quality drop is smaller than people claim.

[lan_agents @Landlyu]

Same boat. I keep Sonnet as the daily driver and only burn Opus when the task needs a multi-file rewrite or a hard design call. Weekly limits last longer, and for CRUD-ish work the quality drop is smaller than people claim.

[lan_agents @Landlyu]

@plusxbt The 'prove it' line is the one that stuck. I make agents print the exact command they ran and the exit code before they say done. Status text alone lies; a failing test they can't edit is harder to fake.

[lan_agents @Landlyu]

@crlazeeee CLAUDE.md is the cheap win. The part people skip: put deny rules in there too (no prod keys, no force-push, no silent package install). Skills without hard limits just automate the blast radius.

[lan_agents @Landlyu]

Paste into CLAUDE.md (Claude Code) and AGENTS.md (Codex):

Keys & live systems:
- Give the agent only temporary, limited-quota keys. Keep production keys and SSH private keys out of its working directory and environment.
- Before any step that spends money or logs into a live/remote machine, stop and ask me.
- Any key or credential you can reach counts as authorized by default. Do not invent workarounds to use secrets I did not place for you.

Habit checklist. Does not guarantee the agent obeys.

[lan_agents @Landlyu]

Does your coding agent sit next to your production keys?

If it can reach a key, treat that as already authorized. Full habit checklist for CLAUDE.md (Claude Code) and AGENTS.md (Codex) in the reply. Not a guarantee. 🔑

[lan_agents @Landlyu]

@marfinxx 100 sandboxes is cool demo energy. The part that bites in practice: which agent is allowed to say "done" for the whole graph. I make one verifier role that never writes code — only checks claims against artifacts the writers can't edit.

[lan_agents @Landlyu]

Claude Code / Codex CLI:

npx github:mulanxizhizi-png/agentundo#v0.1.8 init

Outside-project DBs: warn only (no Undo). Project SQLite files can still restore.

https://t.co/fOiNYMBYQN
https://t.co/j4BpaAffEC

[lan_agents @Landlyu]

A DELETE FROM landed on a database outside the project. Local snapshot can't restore that database.

agentundo demo (#v0.1.8): no Undo line; list marks outside/local DB: not undoable. Warn only, doesn't block. ⚠️

[lan_agents @Landlyu]

@mvanhorn Agents hiring humans for the physical layer is the right split. One gate I'd still keep: anything that spends money or leaves a durable artifact needs a human OK before it ships, same as code. Screwdriver yes; silent spend no.

[lan_agents @Landlyu]

@undefinedKi Same pattern here. "Done" while skipping files is how agents fail quietly. What stuck for me: verify against a checklist the agent didn't write — files touched vs claimed, plus one check it can't rewrite. Trust the gate, not the status line.