Someone at my workplace took 3 days to write a PROMPT to Claude to write a PRD. It generated magnificent 12 notion pages with graphs, schema defenitions, expected behaviors, what-have-you. I read through all of them and raised some questions regrading the architecture he decided to go with, as he went on a different direction that what we discussed, and thats OK. Most of his responses where "oh just ignore that, claude added that". What are we even doing here? The same 3 days could've been spent writing the PRD by hand. Sure, it might not have been so as polished looking as the former, but at least it was accurate! This is exactly my pain with LLMs, they make loads of text, where, they appear efficient. Where actual efficiency is having the least amount of text to describe the feature.

I truly don't know what to do in this area, this goes against the whole premise of writing things up.

LLMs might not be perfect but with these advancements and opus coming into picture, coding as well as documentation have become super easy. I think you should warn the guy of the consequences, PRD/TRDs can be written simply in an hour or two maybe some discussion that requires a day. 3 days is just not done.

Just need to follow some rules for this issue. For example

1. Ask LLMs to be extra concise.

2. Instruct LLMs to avoid diving into details before confirming the general direction.

3. Remove not needed details before asking teammates to discuss/review.

Everything lately is just AI slip specs that people ask me to review. Like, if they spent 10 mins LLMing it I can’t afford to spend 3 hours reviewing it. The same is now happening with code and the quality is decreasing rapidly as we lose critical thinking skills.

Ultimately the trade off for “productivity” IMO is decreased critical thinking for everything offloaded into AI.

just fire them if they can’t think

you have claude for that, to execute your plan

you need to hire a human to make that plan, using AI to improve it not using AI to come up with it. you should be doing that imo.

human thinks. ai executes what i tell it. it can detect patterns and work across dozens of files at once which i cant do as fast.