I can't tell you both how relieved i am to find this article! for two reasons. First, i've been building my own personal OS over the last few weeks and my setup looks very similar to this, so it's reassuring to see that validated in someone else's system. Second, Carl's article has some of the missing pieces of been looking for (: particularly around creating a task backlog. So excited to implement some changes! Thank you so much for this
Carl's approach resonates. The 'infinitely flexible but paralysis-inducing' problem is real.
I went through a similar journey. Started with a simple CLAUDE.md, but after hundreds of sessions it evolved into a multi-tier system - global instructions, project-level rules, and auto-memory that persists across conversations. The structure itself became the product.
Biggest lesson: what you put in CLAUDE.md matters less than having a clear hierarchy. When your agent gets conflicting instructions from three different files, things break in weird ways.
This is super useful. I'm going to try and replicate this on my cloudtop at Google (but with Gemini). I've been able to use a nifty skill which automatically ingests context and the latest detail since last sync of a Google Doc that is attached to a Calendar invite. That way as transcriptions happen I don't know to worry about copying and pasting meeting notes into the CLI; I can just fire up Gemini CLI and in the Gemini.md file it knows to run the scrub of all those documents to ingest context. But I think this tutorial has given me great structure to try and build out as well.
Trying to understand the distinction between workflows and skills. Are skills one-off tasks/operations like "create a Jira ticket?" Whereas workflows start off as a task e.g. "draft a slide deck" but then are further refined by chatting with the agent "rework this section?"
Really like the Projects vs Workflows framing — that alone removes a lot of setup paralysis.
One tweak that helped me: I added a tiny decision rule at the top of CLAUDE.md (fixed repeatable process -> workflow, ambiguous evolving work -> project) so I stop reorganizing folders every week.
Curious if Carl has a cadence for cleaning `_temp/` so useful notes don’t get stranded.
I can't tell you both how relieved i am to find this article! for two reasons. First, i've been building my own personal OS over the last few weeks and my setup looks very similar to this, so it's reassuring to see that validated in someone else's system. Second, Carl's article has some of the missing pieces of been looking for (: particularly around creating a task backlog. So excited to implement some changes! Thank you so much for this
Carl's approach resonates. The 'infinitely flexible but paralysis-inducing' problem is real.
I went through a similar journey. Started with a simple CLAUDE.md, but after hundreds of sessions it evolved into a multi-tier system - global instructions, project-level rules, and auto-memory that persists across conversations. The structure itself became the product.
Biggest lesson: what you put in CLAUDE.md matters less than having a clear hierarchy. When your agent gets conflicting instructions from three different files, things break in weird ways.
Shared my full structure breakdown here: https://thoughts.jock.pl/p/how-i-structure-claude-md-after-1000-sessions
What's been your biggest 'oh I need to add that to the config' moment?
Amazing, thank you for sharing!
@Lily
Ty for rec I saved it
This is super useful. I'm going to try and replicate this on my cloudtop at Google (but with Gemini). I've been able to use a nifty skill which automatically ingests context and the latest detail since last sync of a Google Doc that is attached to a Calendar invite. That way as transcriptions happen I don't know to worry about copying and pasting meeting notes into the CLI; I can just fire up Gemini CLI and in the Gemini.md file it knows to run the scrub of all those documents to ingest context. But I think this tutorial has given me great structure to try and build out as well.
This is great! Thank you for sharing.
Trying to understand the distinction between workflows and skills. Are skills one-off tasks/operations like "create a Jira ticket?" Whereas workflows start off as a task e.g. "draft a slide deck" but then are further refined by chatting with the agent "rework this section?"
If that's correct, I am curious about Aman's take on Carl's workflows vs. skills distinction, in light of reading https://amankhan1.substack.com/p/every-pm-should-be-building-skills
Really like the Projects vs Workflows framing — that alone removes a lot of setup paralysis.
One tweak that helped me: I added a tiny decision rule at the top of CLAUDE.md (fixed repeatable process -> workflow, ambiguous evolving work -> project) so I stop reorganizing folders every week.
Curious if Carl has a cadence for cleaning `_temp/` so useful notes don’t get stranded.