Midwit — Development skill for Claude Code

midwit

A Claude Code skill that stops Claude over-building everything.

Install

npx skills add coopersimson96/midwit

Then run `/midwit` before any scoping question, or say "midwit this" when an answer comes back over-built.

---

Escape the midwit valley.

The problem

Ask an AI a question and the answer lands in the middle of the bell curve. Correct, thorough, well-organised, and three times bigger than the problem. Four phases when one afternoon would do. An abstraction layer serving one caller. A survey of options instead of a decision.

This is not laziness. It is the default reach for legibility: detail reads as rigour, hedging reads as care, structure reads as competence. The middle of the curve is where an answer goes when nobody is charging it rent.

The mechanism

Telling a model to "be simpler" does nothing, because simple is a taste word and taste words get renegotiated by round three. So the skill does not ask for simplicity. It puts a burden of proof on complexity.

**Every element beyond the dumbest working version must carry a `BECAUSE:` line naming a specific thing that breaks without it.** Subject, trigger, consequence: *without X, when Y happens, Z breaks.* No BECAUSE, no survival.

That converts a matter of opinion into a thing that either exists on the page or does not, which is the only kind of rule that survives a long session.

The pass

1. State the objective in one sentence. If it will not close, ask instead.
2. Write the dimwit answer. Dumbest thing that would work.
3. Write the midwit answer. Full size, honestly. Never shown.
4. Diff them. Every extra element becomes a ledger line.
5. Charge each line. Most die.
6. Monk answer = dimwit + survivors. Rebuild, do not edit.
7. Ship it, then list the cuts so anything can come back.

Modes

Mode	Fires on	Ceiling
answer	a question, advice, a judgment call	1 sentence + 3 bullets, 80 words
scope	an idea or project	one page, 5 steps max