Claude Code kept surfacing problems mid-task. I had nowhere to put them.

I'm building Adaptiv on my own, a resistance training app with an engine that adjusts your programme session by session based on what you actually log, rather than handing you a static plan and hoping it still fits in week six. No team, no PM. Just me and Claude Code.

Here's a problem that took me a while to admit was a problem: when you're deep in a build with an AI coding agent, things surface constantly that aren't the task you're working on. Not always bugs, often it's an edge case you hadn't thought through, a dependency that needs updating, a piece of the data model that won't hold up once another feature lands on top of it. I'd be implementing a Stripe webhook and Claude Code would flag that the exercise filter could still serve an overhead press to someone who'd logged a shoulder injury. I'd be mid-refactor and it would surface an onboarding edge case that was going to bite me later.

In a team, someone would Slack it, or drop it in the backlog, or mention it in standup. Solo, with no standup and no one to tell, it just evaporated. I'd either stop what I was doing to deal with it, killing my momentum on the actual task or tell myself I'd remember. I never remembered.

I was losing real, useful observations this way. Not always urgent, but the kind that compound if left unnoted.

The fix: give Claude Code a board, not just a terminal

I set up ClickUp as a proper kanban board and connected it to Claude Code over MCP. Now when Claude Code notices something while it's working on a different task, an edge case, a missing dependency, something half thought-through. It files a card right then, with context, instead of the observation living only in that session's memory (or mine). While it works a task, it comments directly on the card: what it tried, what it found, what's blocking it, so I can glance at the board mid-afternoon and know exactly where things stand without interrupting a session to ask. And on a successful merge, the card closes itself , no end-of-day ritual of reconciling what actually got done.

The auto-open/close part is a nice time-saver, but it's not really the point. The point is I get to operate like a PO/PM over my own build instead of just a developer inside it. I can see what's outstanding, what's in flight, and what's been quietly flagged, without holding it all in my head or hoping I catch it in a commit message three weeks later. For a solo founder, your own working memory is the bottleneck. This just externalises it, with almost no ceremony.

A few things I'd tell someone trying this:

The MCP connection is what makes it frictionless. Claude Code treats the board as just another tool it has access to, not a separate app you have to remember to update.
Define "success" before you turn on auto-close. A merged PR, passing CI, whatever your bar is, pick one, or you'll get false-positive closes on things that aren't actually done.
The board has to be the source of truth, not a duplicate of one. If you're still tracking things in your head as backup, you haven't actually gained anything.

Adaptiv is still pre-launch, so there's no big metric to point to here, just a process change that's genuinely changed how I work. I don't think I'd go back to a plain terminal and a notes app.

Anyone else running an agent against a board like this, Claude Code or otherwise? Curious what your close criteria are, and whether you've hit the false-positive problem.

https://adaptivperformance.co.uk