I wanted to read an academic paper. The paper was long and there were a ton of citations that made reading difficult. I didn’t care about the citations so I wanted them removed. I told Cowork to do that and it did so. Slowly, tediously, and at a cost of 140K tokens.
When it was done I discussed its approach. It turned out that MCP limitations were causing it to perform multiple passes. I proposed doing the jobs in chunks. Cowork agreed and we performed the operation again in chunk mode. That modification resulted in a token savings of 75%. The process still took an annoyingly long time.
While watching it work during this second pass I saw that it was removing citations using a Python script and a regular expression. I asked if that regex could be run in Drafts and Cowork agreed that it could. I therefore triggered my two-day old Design Drafts Action method and Cowork provided the JavaScript containing the regex. I pasted it into a one-step Drafts Action using a script step.
I reran the process. It took Drafts about 2 seconds to clean the 7,500 word document of all references. Tokens expended: zero.
Cowork’s assessment
When a task consists entirely of deterministic transformations (find this pattern, replace with that), the right tool is a script or action that runs locally – not an AI. Claude’s role is to design the script, test it, and register it as a method. After that, Claude steps aside.
For me it obvious. But I’ve seen already so many cases like „Ask AI what is font type and size in PowerPoint template” . AI slowly but consequently takes I(intelligence) from Humans.
@Bjb Thanks for this. You’re highlighting something that is very interesting.
In many cases, MCP isn’t the right tool. As you’ve already pointed out, it wasn’t efficient in this case.
Another challenge with GenAI in general. After every run, you must review the output to make sure it did what you intended. So we pay a cognitive tax for every run.
When the task is a deterministic, as in this case, getting the AI write code is often a better approach.
A question, you mentioned:
therefore triggered my two-day old Design Drafts Action method
My Design Drafts Action is what I call a “method”, David Sparks calls a skill and others call a saved action. I discussed the process design with Cowork and it wrote the method.
There is another side to the efficiency story. Cowork estimates that reading, writing, and processing drafts using the Drafts MCP imposes a 240% efficiency penalty over the same operations performed directly on plain text (Markdown) files. An operation performed directly on a file that costs 1,000 tokens costs about 2,400 tokens if you use the MCP.
That is not as bad as the penalty for dealing with more structured data through MCPs provided by Tana and Notion. The penalty for those MCPs increases to about 500%.
I’m adjusting my Cowork ←→ Drafts interactions to account for these differences. The example in the initial post describes a situation where an operation in Drafts is a much better choice. But, if you direct Cowork to use the MCP to read and manipulate a draft that is much less efficient than using a Drafts Action to create a file that Cowork can work with directly.
. AI slowly but consequently takes I(intelligence) from Humans. 