The Handoff is Dead

The back-and-forth is a yesterday problem.

For months I’ve been asking Claude what it remembered, or remembering on its behalf (yes, you can tell it to remember specific things, I know). I dictated specs. I copied output back into documents. I wrote prompts I’d written ten times before. Claude was working for me… and even though I could have written skills for some of it, I was still doing a lot of work.

I rebuilt my workspace. Last week. Couple hours.

Same surface for Claude and me. Same files. Same instant. My backlog lives in Obsidian. Claude reads it. When I edit a spec, Claude sees the edit. When Claude writes a security report, it drops into the vault and I watch it appear. Altogether, in real-time. I don’t have to ask Claude what’s what anymore! Now I see what Claude sees… now we’re working together vs. Claude working for me.

Here’s what I learned. Your AI assistant is hungry, and you are tired. Markdown is a buffet at which AI likes to eat. Feed your AI what it likes to eat.

Credit is due: Andrej Karpathy has been beating this drum on X around LLM Knowledge Bases. Not my idea. I just picked the simplest possible frame so anyone can copy it. Cost: $4/month for Obsidian Sync.

So what? Chat is for steering. The vault is where the work lands.

That shift unlocked the next one.

Remember the 6 prompts I gave you last week to audit your app? I converted every one of them into a Claude Code skill. /security-audit. /hostile-user. /ux-expert. Single-line invocations, timestamped reports, resolved reports move to archive. The audit trail builds itself.

One-shot prompts are training wheels. You type the same thing every time, hope the model is paying attention, paste the output somewhere if you remember to. The work evaporates the moment the chat ends. Skills are scaffolding. Scope, severity rubrics, output formats. Same review every time, on demand, with a paper trail. They live in the vault, right next to the specs and the backlog.

That’s the loop.

Documentation rots because it’s not part of the work. The fix is making it part of the work. When a TestFlight round closes, the changelog updates. When I ship a schema migration, the spec updates. Altogether, now!

Persistent context. Compounding documentation. Audit trails for every review. Switch laptops, the vault is already there, the conversation resumes (need I say, the buffet is open!).

Most AI productivity content is about one chat session. This is the infrastructure between chats. The setup isn’t optimized. Optimization happens project by project. But it works – right now.

I attached the full layout. Folder structure, sync mechanism, doc-maintenance triggers, the skill pattern. Steal what’s useful. Hell, feed this to your AI and ask for help setting it up.

The handoff was the bottleneck. Kill the handoff.