Punderstruck

The pun skill your AI agent didn't know it needed.
Real-time phonetic discovery. Comedy theory. A 17/25 quality bar.
No recycled jokes. No "lettuce" puns. No mercy.

---

The Problem

Ask any AI to make a pun. Go ahead. You'll get one of three things:

1. A pun from 2019 that your uncle already forwarded in a family group chat
2. A pun so obvious the audience can finish it before you do
3. Something that isn't actually a pun but the AI is very proud of it anyway

This happens because LLMs generate puns from training data. They're pulling from memory, not discovering new connections. The result is the comedy equivalent of reheated leftovers.

What Punderstruck Does Differently

Punderstruck doesn't retrieve puns. It **discovers** them.

The skill queries the [Datamuse API](https://www.datamuse.com/api/) in real time for phonetic collisions (sounds-like, homophones, compound splits), then builds puns from scratch using structured comedy construction (punchline-first method). Every candidate gets scored against five quality criteria. Only puns scoring 17/25 or higher get served.

The entire pipeline runs silently behind tool calls. You see the pun. Nothing else.

**Generic AI pun** (what you're used to):

Why did the AI go to therapy? Because it had too many issues with its layers!

**Punderstruck** (what this skill produces):

I asked my AI to lose some parameters, but it told me to weight.

One was pulled from training data. The other was built from a Datamuse collision between "weight" and "wait" applied to the neural network domain. That's the difference between retrieval and discovery.

Quick Start

# Claude Code (recommended, auto-upda