Every Claude Code session starts the same way. You say “run the tests.” The agent says “I’ll look for how to run tests in this project.”

5 points•runtooldev•4 days ago•2 comments•

2 comments

runtooldev4 days ago
There was a comment about just using AGENTS.md instead that got deleted, so I'll answer it anyway:

Natural language markdown instructions are prone to hallucinations or picking the wrong flags, whereas MCP provides a rigid, typed schema with zero guessing. I guess it has to be experienced to be believed, but the speed with which agents navigate project scripts and database queries with the right MCP tools exposed is insane.

verdverm4 days ago
I think that freedom has its benefits

saw our new review agent comment on a line not in the diff, it pointed at the actual RCA the pull request missed

they also pick up local conventions by looking at the files around, the example in this post is rather contrived, run tests, and doesn't align with the save exploration tokens claim

the main problem I've seen with MCP is that they become a dumping ground and you end up with a ton of context every session from N descriptions, 3rd party vendors are netoriously bad at these and skills, gotta jam every feature we have in there!

All that being said, I think I might try to replicate something like this as a plugin to OpenCode. Today I use markdown and makefile for consistent easy commands. A small config that looks like a tool, not mcp, has me intrigued

Read the full thread on Hacker News →

Related stories