Claude Code CLI Skill For Pickaxe

Community Scripts
1

I built 80+ Pickaxes with Claude Code helping me, and the Claude Code kept making the same expensive mistakes. It would edit prompts on live tools, because Pickaxe has no draft mode. It would report a fix as done when the platform had silently discarded the write. It once “fixed” 80 things that were never broken.

So I packaged everything my assistant learned the hard way into a free, open source skill:

A skill is a folder of instructions that AI coding assistants read on their own. This one works with Claude Code, Codex, Cursor, Grok Build, and Grok Bot. Install takes one command, and there is nothing to configure.

Once installed, your assistant arrives already knowing things like:

  • Every prompt write hits your live bot, so it tests on a staging copy first
  • Input caps can silently feed your bot one page of a 300-page manuscript
  • Adding a document to your workspace does not attach it to your bot
  • Long runs die at exactly 300 seconds, and retrying never helps
  • Replacing an action’s manifest silently resets your trigger prompts

It also walks you through connecting your assistant to Pickaxe’s MCP server for the first time, including where the API key hides in Settings. And it teaches your assistant to interview you about what a good answer looks like before testing your tools, so it can judge results instead of guessing.

Everything in it was observed on the live platform and carries the date it was observed. It is an independent community project, not an official Pickaxe product.

If you try it, I would love feedback. Reply here with what worked and what did not, or open an issue on the repo. And if you hit a platform quirk the skill does not cover yet, that is exactly the kind of contribution that makes it better for everyone.

2

Update: v0.3.1 is out, with about thirty new lessons since the launch post

Two releases have gone up since I posted this. Both came out of running the skill against real work on our tools, and a few of the additions came from things the skill got wrong the first time.

A new reference file for judging output. The testing file was getting long, so grading now has its own reference. It covers how to write a rubric before you test, how to build test inputs that can actually fail (a deliberately bad input paired with a good one, so you can tell a tool that improves copy from one that just rewrites everything), how many runs an A/B needs before a percentage means anything, and how to grade a tool that quotes your document back to you. If you have ever eyeballed three outputs and called a prompt change a win, this file is for you.

Corrections to things the skill said at launch. Three of the launch claims turned out to be incomplete, and they are fixed. The rule about raising a field’s cap and its type together applies to text fields only, since upload fields have no type to move. The list of fields the API silently ignores is per field, not “display labels”: descriptions save fine, the submit button label does not, and both can sit in one call. And the workspace fallback model can change your reasoning effort as well as your model, which matters if you tuned a tool at a specific level.

Platform facts that were expensive to learn. A few that stand out:

  • A new Pickaxe created over the API arrives public, with an upload budget that cannot work and a 250 token chat cap. The skill now lists what to overwrite before the first test run.
  • Toggling a user’s access group in the admin drawer disables their current group and wipes any credits they purchased, with no undo. Pickaxe confirmed this is by design. The skill now says to use the API path for group changes and to stay out of the drawer for paying users.
  • A truncated upload does not just make output thin. It makes the model invent things to fill the gap. If a user says a tool “made something up,” check the upload cap before touching the prompt.
  • Image actions return public files on Pickaxe’s CDN. A user’s “access denied” on a picture is usually the portal’s lockout page, not the file.
  • The public completions API has now been exercised with streaming, form field injection, and image URLs, and a slow tool that kept dying at the 300 second ceiling over the MCP tool finished six times in a row with streaming on.

The bundled Python client got better. It can now drive a form tool through its real form fields, not only through a chat message, which turns out to be a different code path on the platform. Testing that fix also caught that the platform changed the shape of completion responses between August and September, so the client now handles both.

The skill description is measured now. I ran it through a trigger test with twenty realistic prompts, half of them near-misses like the npm package named pickaxe and Minecraft. The new wording fires on 7 of 8 held-out Pickaxe prompts and never on the decoys. The prompts are in the repo under evals/ if you want to run them against your own edits.

Same install as before, and if you already have it, git pull in the skill folder gets you current. Release notes for both versions are on the GitHub releases page. Still an independent community project, and still very happy to hear what breaks.