Error propagation

error propagation is a real issue with agentic development.

if you've been prompting and editing files for hours, there is a high chance that at one point you'd just type "sure go on" after skimming through an agent response. typically this shouldn't be an issue, BUT if one fact is incorrect in that response then you'd have to suffer its fruits for the rest of that session. (until you correct it of course).

that's the typical reason people report that "claude isn't doing what i told it to". it's mostly because it has a wrong context that you are trying to out-prompt instead of correcting. but you can't correct what you don't read.

no matter what Anthropic's marketing team tells you, the "memory" feature they sell us isn't memory as we humans experience it. it, at its best, just pastes saved notes into the context before the model reads it.

models don't have memory. zero recollection in static. in simple terms, at every junction (new prompt from you) they gather your past prompts AND their own responses in that session, and use that as the foundation for their next response.

so a wrong fact you didn't check will be there AT EVERY TURN.

wrong solution: (take it with a grain of salt, i am by no means an expert at this *19) ==> i used to think asking agents to respond in shorter sentences makes it easier for me to spot issues. but the problem with that is it just summarizes! so the error gets summarized into the response, and summarizing a factual error doesn't make it less of an error (that was a mouthful. pause).

my solution: (see *19)

don't ask for shorter responses. ask it to restate the plan before it touches anything. not the long winded version, just the plan. steps, in order, one after the other.

also take some time to expand the "thinking" dropdown. see what the ai is thinking, cause that too is context. and if you can spot a possible issue fast, all the best!