Here is a major context engineering bug in Claude Code that bloats output token usage.
Have you noticed the below? Tell us about it!
Observation:
Claude Code agents often repeat verbatim file contents in their subagent prompts.
Why that is a problem:
It could have provided file paths for subagents to read. Instead it wastes output tokens, which are much more expensive than input tokens! Imagine that happening for "n" parallel subagents being invoked.
Why this may be happening:
- Deliberate instructions in system prompt to do so: I cannot imagine a reason for them to do this, except higher billing.
- Bad judgement by models: Which could happen due to confusing instructions in the system prompt, or due to lack of guardrails for this scenario of agent invocations.
This fabulously written article by @mitsuhiko about the perils of humanizing AI gives me a very MMORPG vibe about our landscape!
A lot of the discourse around the word is opinionated expressions, which is totally fine. The important bit getting lost in the discourse is at the end of the article - the boundaries and the responsibility.
This is why data centers are coming to India because they are facing this push back back home. Here the people and the government don't care if the water is that color or poisonous as long as they can hate and discriminate against minorities.
One of the underrated reasons Linear is so popular with so many people is they have an internal target that nothing in their interface should take more than 300ms to render. They keep fixing regressions whenever it happens.
It’s very hard to retrofit this culture: look at JIRA…
I'm not sure if knowing/realising that makes much of a difference. It's frustrating floating on the surface while the root cause is trying clean it up.
The conjecture does indicate that maybe it may not lead to anything mainstream, but that won't stop pent up energy from taking its natural course.
Also, the Gen Z populace comprises more than just the exam takers.
This is huge trouble. 2 months back, I had to haggle with Urban Company's AI and then with the human(eventually!) for cancelling a mis-scheduled appointment. Sadly, it ate up almost half an hour of my time.
Anyone knows how to reach a human at @makemytrip ? Their AI chatbot does not understand my problem. If I ask to talk to a human it gives me a phone number that has another AI agent with speech recognition. None of them understand the problem. Here I am stuck with Rs 10k gone.
Anyone knows how to reach a human at @makemytrip ? Their AI chatbot does not understand my problem. If I ask to talk to a human it gives me a phone number that has another AI agent with speech recognition. None of them understand the problem. Here I am stuck with Rs 10k gone.
That is a necessary callout!
Not that it is less technical, but I am planning to try out @openfangg.
For the record, does anyone here know alternatives that are great UX for non technical people?
Dont say Claude Cowork!
1. It is not "great" UX. It is broken in many places.
2. I am talking about alternatives that are closer in capabilities when compared to the (generally)claw-league.
@GergelyOrosz I usually start with "I'm considering approach X to achieve Y. What are my alternatives and which one do you think is a balanced and correct choice? Point out pros, cons and tradeoffs for each. Be critical and concise."