Experiences & Lessons
What went wrong, what to avoid and what actually worked. Written after the fact, from real use, not from the docs.·Page 2 of 2
Automation agents most need stable execution and deterministic output, and OpenClaw falls short on both: the same task gives different results, failures go unreported, and success cannot be guaranteed. Our fix is to add inspection logs and failure notifications, script the processing steps, templatize prompts, and make outputs JSON with strict validation.
Feedback from learner Feiwen: before, every problem meant searching everywhere, costing hours in total; now, opening an LLM clears up doubts in three minutes, and learning is much faster. We break down why it's faster, and how to use it to be both fast and accurate: give full context, ask one thing at a time, ask about the principle behind it, and verify the key points.
An experience piece (members only): for occasional use (chat, copywriting, code edits), an API is cheaper; for high-frequency batch work (daily transcription, classification, processing large volumes of material, fixed workflows), running locally may pay off. The core dividing line is monthly token consumption, plus whether the task needs top-tier model quality. The article works out the numbers and the hidden costs on both sides for you.
A practical lesson: models confidently present wrong things as right. Here is a prompt you can paste directly, which forces it to weigh evidence for and against based on authoritative literature, news and public discussion, give sources, separate facts from speculation, and state what information is still missing. Usage note: many models do not actually browse the web, so check the sources yourself.
Some source code can pass Cloudflare's human verification, but models like Claude have too strong a moral stance and refuse to help you get around it. Consider using another model.
Hands-on test on November 6, 2026: after opening a few web pages and running a few extremely simple tasks, the Grok bot quota was nearly used up within two days, with 35% already consumed. After the update at 5 PM on November 5, the workspace kept failing to open. Feedback has been sent to X's official team, with no response so far.
A learner handed in a Claude draft almost untouched and got an academic-integrity hearing three days later. What gave it away, what it cost, and how to actually use AI when you write.