A small repo of AI skills for product managers. You clone it, you run it in Claude Code, and it is yours — including commercially, including if you never speak to me again.
git clone https://github.com/aragjha/prodman-stack
cd prodman-stack
claude
# then, in Claude Code:
> Read skills/portfolio-from-work/SKILL.md and run it on my last project.It will interview you. Answer from memory — you do not need to find the documents. That is the point: your memory of a decision is not your employer’s property, and it is the part that was always the interesting bit.
Nothing leaves the folder carrying your employer’s name, your customers’ names, or a number you are not free to publish. The skill refuses to write a case study until the de-identification pass has run, and it shows you exactly what it stripped so you can check it rather than trust it.
prodman-stack/ ├── skills/ │ ├── portfolio-from-work/SKILL.md ← work you can't show → work you can send │ └── launch-minimum/SKILL.md ← the eight things, and nothing else ├── evals/ │ ├── golden-sets/ ← 5 adversarial inputs per skill, traps documented │ └── scoreboards/ ← results, failures left on them ├── examples/worked-example.md ← one real case study, start to finish ├── experiments/LOG.md ← every hook tried, and how it did └── LICENSE ← MIT. Use it commercially.
One skill that works beats forty you never run. More get added as they prove themselves on real work — not before.
Not bugs — the difference between a folder of skills and a system. The repo is genuinely yours under MIT either way, and if the one-shot is all you ever need, that is a completely fine place to stop.
One new skill lands in the repo as it is built and tested, each with the eval golden set it has to survive — five frozen adversarial inputs with every trap written down, so you can audit the checks instead of trusting them. Drop your email and you get each one the day it ships, along with the scoreboards as runs happen — failures included.
No newsletter. No drip sequence. A skill, when there is a skill.