Superproject, ShipCheck and the Other Tools I Built for Claude Code

I work with Claude Code every day. Not as a toy, but as a tool that generates production code. (More about me and my work with AI.) And when you use something this intensively, you quickly hit edges that slow you down.

Instead of complaining, I wrote my own solutions. Most stayed small, fixing one annoyance each. Two grew into something bigger.

Superproject: a project that moves on its own, but never without you

The problem first. Claude Code is excellent at a single issue. It gets harder the moment you have more projects than hands and the only thing you genuinely have time for is review. Briefs fall apart, the backlog ages, and you spend evenings writing instructions instead of making decisions.

Superproject is the layer that closes that loop. It takes a brief and turns it into a backlog with testable criteria. Vague briefs get rejected with clarifying questions. The backlog is pushed to GitLab or GitHub as issues. Then it takes one issue at a time: plans it, has the plan adversarially reviewed, implements it, tests it and opens a merge request.

And here is the part that matters.

The package ships ten skills, from initial planning through platform sync and autonomous runs to maintenance, replying to issue comments, and closing the project out against the original brief.

It sells as a one-time licence with twelve months of updates, not a subscription. Current pricing and documentation live at superproject.dev.

ShipCheck: nine questions before you ship

The second product came out of a different worry. Vibecoding can produce an app over a weekend that looks finished. The question is not whether it works. The question is what happens once a stranger sees it.

ShipCheck takes a deployed app and runs nine checks, phrased in plain language rather than security jargon:

1. Who can read your data
2. What a second account sees
3. Whose bill this lands on
4. Whether the money path holds
5. Whether anyone can find you
6. Whether your AI can be tricked
7. Whether someone can break into an account
8. Whether the app trusts the browser
9. Whether you are covered legally

The result is a three-level verdict: do not ship, ship once the fixes are merged, or you are fine. Every finding is proven against the running app rather than argued in theory, and fixes are tried on a throwaway preview build. Nothing deploys automatically.

ShipCheck is included in Superproject as the final gate before launch. It is also sold on its own at getshipcheck.com.

And then six small things that stayed free

These were never meant to be products. Each came out of one specific thing that stopped being tolerable, and all of them are open source.

claude-replan is the most important of the lot. Claude Code can write a plan that looks great, and the flaw surfaces halfway through implementation. /replan sends several agents at a finished plan at once, each from a different angle, and reports what does not hold. /recheck does the same after implementation. It has repeatedly shown me a check reporting success without testing anything. The full story is in a separate article.

claude-gc cleans up orphaned processes. Claude Code spawns subprocesses, subagents and MCP servers, and when you close the terminal some of them hang around. Each eats around 220 MB. I have seen over a thousand orphans consume 75 GB on a 32 GB server and take it down with an OOM kill. The tool walks the parent chain, so it never kills your active session. Pure Bash, zero dependencies, runs under cron every 15 minutes.

claude-docu-optimizer deals with CLAUDE.md files that bloat over time and stop matching reality. Claude then works from inaccurate instructions and you cannot tell why. The tool detects ten kinds of anti-pattern, from verbose content that burns tokens, through documentation that no longer matches the code, to files nothing links to.

claude-notify-telegram is the simplest of the set. You start Claude Code on a long task, walk away, and have no idea when it finished. The plugin hooks into the events and sends a Telegram push. With no credentials configured it silently does nothing and blocks nothing.

remoteclaud turns a bare Rocky Linux ARM64 server into a complete development environment in one command. An Ansible playbook that installs Docker, Node.js, Python, Go, the Claude CLI and a tmux session manager. Type cl and you get a menu of your sessions.

DropShot is the only one not directly about Claude Code, though it was built with its help. A native macOS menu bar app: drop a file on the icon, it uploads over SFTP and the path lands in your clipboard. In a tmux session over SSH there is no drag and drop. I built it in one afternoon from my phone, and there is a separate article about that too.

A complete overview is in the Tools section.

Why I do this

None of these projects were planned as products. I was solving a problem that slowed me down, then noticed that everyone working intensively with Claude Code has the same problem. With two of them it turned out the solution was worth taking further than a script for personal use.

That, incidentally, is the main reason I care about this work. I do not talk about AI from a stage. I use it on my own projects and I carry the consequences when something breaks.

Want something like this in your company?

Most of these did not start with me sitting down to write a specification. They came out of ordinary work, where something annoyed me and AI helped me fix it over an afternoon. You do not need to be able to build products for this way of working to make sense for you.

That is exactly what we do at company workshops. We take one thing that repeats every week at your place, and you leave with a process that still works on Monday morning. Seven people maximum, NDA as the first step, work on your real material, not on demo examples.

Take a look at the workshops or just get in touch.

You might also like

• How I connected Plaud to OpenClaw and got better transcripts
• Why I never read the first plan from Claude Code
• DropShot: a macOS app in one afternoon