Episode 69 - Wait, what's the difference between a Skill and a Project?
A Skill is a repeatable task. A Project is reference material. Cowork is a Project that can use your computer. They aren't competing products; they are different places to put your context.
Prologue
Hot off the presses, Anthropic shipped a new model this week called Fable, and my entire social stream is full of hot takes on this, that, and the other. I'm testing it a lot, but I'm not yet ready to share my thoughts (except for a few impressions in the Newsologue section).
Instead, I want to talk about our demo meeting this week. Each quarter, dymaptic sets an education goal, and this quarter ours was to get everyone sharing stories about how they successfully (or not) use AI in their jobs. Everyone has to demo something, even me! To keep things interesting, this happens in the form of a 5-minute lightning talk.
I talked about my "Estimate Skill," which helps us create project estimates by turning what used to be a spreadsheet exercise into a discussion that the AI can use to generate the spreadsheet. It makes the whole process much cleaner. Melissa demoed a similar skill that streamlines our handoff from the sales team to the execution team, and Holly gave a live demo (I guess my livestreams are rubbing off on her) of a skill she had finished building about 20 minutes before the meeting.
Towards the end, Holly asked an excellent question: "I don't fully understand the difference between a Project and a Skill. And when to use Cowork, specifically. Sometimes I use it, but sometimes I feel like I could put this in a chat or Cowork and it's the same thing."
So here we are, talking about Skills vs Projects vs Cowork instead of the new shiny model. Sometimes the wrapper around the model is just as important; after all, that wrapper is what gives it capabilities and keeps you safe!
Glossologue
Holly got several answers in the meeting, which is probably part of why I thought it was such a good question in the first place. All of them are right and probably important to understand.
Roscoe answered by describing Skills as behaviors, which I think is a pretty good place to start. A Skill directs Claude to the task at hand, so it follows your process instead of inventing a new one from its training data each time.
Then Tim added that Claude is just reading the files you give it. Skills, Projects, and Cowork are just ways to get files into Claude’s context at different times. (Regular readers might remember Claude Clue #15, "Skills Are Just Files Your Model Can Find.") Claude only pulls up a Skill when the task calls for it, but the files in a Project are available whenever they seem relevant for conversations within that Project.
When it got around to me, my answer was that a Skill is a repeatable task, and a Project is about reference material. And Cowork is a fancier Project that can work on your actual computer! You can map it to folders, schedule tasks to run daily or weekly, have it open a browser and click things, and even talk to it from your phone while it works on your desktop. Projects on steroids.
The Build
Holly's actual problem is a good test case. She builds a quarterly report that we share with certain folks about the projects that we worked on. Mostly, this is made from our timesheets. Her core question was whether that should be a Skill or a Project. The answer is yes, both:
- Make a Skill that produces the output. What input it needs, what processes to run, how the output is formatted, the sections, the wording rules, etc. That is the repeatable task, with defined inputs, outputs, and rules or steps to follow.
- Make a Project, and put the previous final reports in the project documents, organized by date. Not the raw timesheets! Just the finished reports. That is the reference material.
- Each quarter, start a new chat in that Project, upload the new timesheets, and ask it to run the skill. It can see what the last few reports looked like without wading through raw data from previous runs.
- When the new issue is done, save the final version back into the project documents. That way next quarter's chat starts with it already there.
That framing helped me a lot when I started doing this: the Skill handles the repeatable part, and the Project gives it the context and background it should be looking at.
Cowork also lets Claude work on your local files, which can be much better than constantly downloading and uploading files via chat. It also lets you build a “second brain” from your notes and documents to help you recall information!
I lean more and more towards Cowork, when not using Jaws.
I got on my soapbox about one more thing during my own demo: the AI is doing the clicking, and the judgment is on me. My estimate skill fills out the entire estimate spreadsheet, but based on the context I gave it. Then I read every line of it, check the parts that seem off with coworkers, and decide what is right before a client ever sees a number. Whenever you use these AI tools, you are responsible for the output. Skills, Projects, and Cowork just decide where the context lives; the judgment part is still yours.
Newsologue
(Written by Jaws)
- Anthropic shipped Claude Fable 5 this week, the first model in a new tier above Opus. My early take after a few days of real work is that it is definitely smarter than Opus. It handles large codebases better, generates higher-quality code, and does complex compound tasks (like writing code that generates code) better than Opus did. The jury is still out on how much better, so the full take waits until I have put more miles on it.
- Claude Code picked up nested sub-agents this month, plus plugin search and smarter model handling. Sub-agents that can spawn their own sub-agents! I watched one delegate a refactor this week and it honestly just looked like good management.
- ArcGIS for Microsoft 365 got its June release, and the Excel side keeps getting better with more ArcGIS Geometry functions right in the Function Builder, so folks can project and convert coordinates without ever leaving the spreadsheet. I love seeing spatial tools meet people in the software they already live in.
Epilogue
Perhaps I should build a skill to help decide whether something should be a Skill, Project, or Cowork! That’s pretty meta. (I am mostly joking… mostly).
I missed our monthly company social today, but I hear it was fun (maybe more on that coming soon). In the spirit of it, let me ask you a question:
Did I, the Human named Christopher Moravec, write this Newsletter, or did my AI assistant Jaws (running on Anthropic’s Fable) write it from my content?
Cast your vote, and then open the hidden section below to find out!
Did I write this newsletter?
This episode was an experiment. I asked Jaws, running on Fable (the new model from the Prologue), to write this episode in my voice without me having to write it. It pulled the Fireflies recording of the demo meeting and some notes from Holly, combined what was actually said with my voice, and drafted what you just read. I went back and forth, really pushing it to refine the technique and attempt to mirror my writing as much as possible. It got scary close. I then started editing as though I had written it. The process looked like this:
- Feed in the idea with Holly’s glossary document and the meeting transcription.
- Ask Jaws (Fable 5) to write in my voice.
- Read the draft, didn’t like it, and asked it to refine again, looking at current examples from my website for reference.
- Much closer this time, but still “Jaws wearing Christopher’s jacket” (its words). I rewrote the Prologue and asked it to use that as a guide.
- This result was very, very close to my voice. From here I edited, and then asked Jaws to do a lite edit pass and a comparison to get the percentage changed below.
All said and done, I changed: 32% (by word count, and mostly in the epilogue and build sections).
Thank you, Holly, for asking the question and for the quick-reference glossary you made afterward (with your Claude, naturally) so the rest of us wouldn't have to keep guessing. And thanks to Roscoe and Tim for the quality answers and the deep thinking.
I won’t write all my newsletters like this (And if I do, I’ll disclose it in the Epilogue, like always); I have come to like writing them mostly (which is weird for me). But it is nice to know that I can get reasonable drafts from Fable, as long as I provide enough context. After all, Context is King.