In short
A bad result is data about what the model understood. Read it, adjust, go again. Beginners stop at two tries. Builders go fifteen.
The first practice. When a model gives you something wrong, it is telling you exactly where your instruction and its understanding diverged. The wrong image, the wrong tone, the missed rule: each is a precise pointer to the sentence you need to change.
People who get good at working with AI treat every output this way. They do not judge the model. They read the gap and close it. Fifteen rounds on a hard problem is normal. Two rounds and giving up is the most common failure we see.
The same habit applies to a running agent. When it does something wrong, the fix is a sentence in the instruction file, and the sentence usually starts with the exact mistake.
Example
What it looks like.
Round 1: the post has two images stacked. Round 2: add 'never two image components adjacent.' Round 3: correct. The rule now lives in the validator so it never comes back.Related
Read next.
- promptWhat you send a model to get something back. In a chat, a message. In an agent, the instruction file plus the situation. Save them. Improve them. Never retype them.Read the definition
- validatorA script that refuses bad output before it ships. The agent does not get to decide whether its work is good enough. The validator does.Read the definition
- program in txt and mdThe instructions are the code. Write what you want in plain language, keep it in a file, update the file forever. No Python, no Bash.Read the definition