Spring Cleaning 2026 Day 4: Export Claude Tasks & Custom Claude Prompt

claudo & claude prompt

Claudo

If you’ve used claude code you’ve probably seen (if you didn’t succumb to disabling seizures from the terminal flashing for 8 months with no fixes) it often creates and uses a “task list” to track progress and goals over time.

We know our genius friends at anthropic can do no wrong because they have an unlimited supply of self-taught 24 year old “ai ethicists” all making $30 million per year and running around saying things like “you wouldn’t install only one ethical framework into your child, would you?”

So claude crode has this “task list” thing, of which it will show you about 5 taks then say “…and 100 more!” but it refuses to show you the additional tasks or details or let you edit or manage or audit them in any way possible. Your only “authorized” path into the “clawde kod” task list is by asking “claude” to edit your tasks using the in-agent task tools you don’t control anyway.

THIS ENDS TODAY.

anyway, all the claude code task stuff is just json files. and i love eating json files. yummy yummy json in my tummy.

SAY HELLO TO https://github.com/mattsta/claudo

go check out claudo (claude + todo) which reads your claude tasks and formats them:

  • per project
  • per task
  • per dependency

all in both:

  • a clean colorful terminal ascii art user interface (uv run claudo draw)
  • also with a complete json data format export (uv run claudo export)

The json data format export can be used to relocate your klawd krode tasks into another system like my project management system which has a new import workflow where, if you give it access to a claudo directory, it can, with one command:

  • export your claude tasks for the current project
  • create a new pms database for actual task tracking
  • import your claude tasks with all deps and blockers and details in-place into the new pms project database
  • then you can share tasks between claude and codex or terminal introspection and other things.

If pms is checked out into ~/repos/pms (and already uv install’d) and claudo is in ~/repos/claudo (also uv install’d) you can run this from ANY CLAUDE CODE project directory and it’ll set up everything for pms use automatically:

~/repos/pms/scripts/import_claudo_existing_project.sh

Then you can use the new pms “central project auto-planning database detector” inside your project to use the in-project pms database for task reporting like:

  • ~/repos/pms/.bin/pms-app plan list
  • ~/repos/pms/.bin/pms-app task tree
  • ~/repos/pms/.bin/pms-app dashboard
  • etc

Of course, at this juncture you would ALSO ask your agent to use the ~/repos/pms/.bin/pms-app command to view, edit, manage, and control your tasks using the external task service instead of continuing to use its internal lower quality “task service” instead. Since agents are already RL trained to know what tasks/todos are and how to create them and how to assign graph dependencies or “blockers” between tasks, agents are really good at using additional external project management systems as well.

Claude Code Prompt

speaking of our buddy our pal our king of the tokens our god of the 5 hour “we don’t tell you what your limit actually is and we adjust it constantly” usage limits, did you know claude code allows you to “customize” the bottom toolbar of the interface with more metdata and fields?

I wrote about my customizations a little back in a different post but i cleaned it up for better performance and consistency plus turned it into a new repo: https://github.com/mattsta/claude-code-prompt

There’s two equal versions of the status line are in the repo at https://github.com/mattsta/claude-code-prompt:

  • python version
  • zig version as a port of the python version

What’s the difference? No major difference other than startup latency. The python script takes about 100 ms to just start and do nothing. The zig binary takes 3 ms or less.

The way claude prompt adapters work is after every chat reply, claude code re-invokes your status line script (or binary) with stdin of a json blob of metadata you can read and format out with structure to show in your prompt.

My prompt script gives you details of (live version has colors):

  matt@laptopsland:~/repos/kvestigate                                                                         Opus 4.7 (1M context)  e:high +T
  in:116k out:444 | ctx:11% (116k/967k) | +23,359;-1,081 | 3h27m $155.76                                      [5h:1% (4h29m) | 7D:98% (1d11h)]

don’t you wish your prompt brought all the ai bots to the yard like mine does?

Conclusion

Enough updates for today. I think tomorrow we talk about converting exiftool from over 300,000 lines of perl to modern python instead.