The other day I had two hours of meetings in the morning. The first was standup. Earlier I had created 10 subtasks under one parent task. I was asked to convert it to an Epic. I went to Claude Code and asked Opus 4.8 to move them (overkill I know but I normally have spare credits and I've been getting better overall usage using a better model with fewer iterations than a worse model and more).
I have a couple of skills, one that prepares a work item for work (checks that it meets a definition of ready etc) and another that starts work on it. They might each read an item and update it a small number of times. I ran each of these skills one time.
My meetings ended and I went to do some actual work. Only a few minutes in and I got the notification in Claude Code that I had used 99% of my 5 hours usage. Which was rather shocking. I very rarely hit the 5 hours limit except with intentional heavy usage and here I had been in meetings all morning and barely used it.
`/usage` reported that in the last 24 hours:
35% of your usage came from MCP server "plugin:atlassian:atlassian"
In the past I had used Copilot CLI with the MCP server and that would very frequently warn about large responses whenever I asked it to do anything with Jira. These weren't even large items.
I am trying to switch all my skills to use the Atlassian CLI instead. Unfortunately that doesn't support Confluence page writes yet (please add that).
I'm not one of those people who say that MCP is dead (that's a stupid argument, bad harnesses including all tool definitions were the real issue). That being said, you can save a lot of context and round trips to the model in custom skills by using custom scripts that glue together a bunch of CLIs.
Confluence page updates are very painful, especially on moderate to large pages. I haven't looked to see exactly what is going on but even Opus frequently gets the ADF wrong and has to try multiple times. I even saw it try and split the page into parts and do multiple requests which it did in a really odd way and it had to keep checking to verify that it had stitched it back together correctly. It managed it in the end but it seems way harder than it needed to be. This is especially problematic for small edits.
It would be really nice if you could:
a) Minimise context usage across all the tools
b) Better support for small edits of confluence pages
c) Better support for edits of large confluence pages
d) Add confluence page edits to the AWS CLI (I know this group is about MCP but as a user they are all part of the same toolkit)
