Prompt lab

Most prompts that get shared were never run. They were graded, if at all, against rules someone remembered, and prompting advice moves every time a model ships. So the prompt reads well, then falls over the first time a stranger pastes it into a fresh chat.

This skill writes one prompt properly: graded against guidance fetched this turn, held to a bar a non-developer can run, and run by Claude before you see it. It makes one prompt. It does not package a whole process for an audience (that is workflow-maker).

Hard rules

1. The finished prompt comes back in the same turn, in a code block. Never a template to fill in, never a command for the user to run. Ask a question only when two readings of the idea would produce materially different prompts. Otherwise state your assumption in one line and write it.
2. Fetch Anthropic's guidance live, every run. Never grade from memory or from a list baked into this skill. Techniques fall in and out of favour between model releases, so a remembered rubric becomes the stale advice it was meant to catch. If URL 1 fails, the Anthropic half is NOT RUN, and there is no total out of 100. If only URL 3 fails, grade from URL 1 and name the URL that failed.
3. Every point lost has a reason. A score with no evidence behind it is a guess.
4. Run it before you hand it over, whenever you can. If the prompt needs nothing you lack (a file, a connector, a paid tool, a person), run it on a realistic sample input and check the output against every finish line. If you cannot run it, write TEST RUN: NOT RUN and the reason.
5. The pass mark is 95. Below it, fix and re-grade. Never round a 93 up, and never call a prompt finished that scored under 95.
6. Quote, never invent. Every Anthropic row quotes a sentence from a page you fetched in this run.

Step 1: Pin the job

From the idea, settle four things. Infer them. Do not interview the user.