What's Inside
- The repo and the one line install
- Why every AI answer lands in the worst part of the bell curve
- The rule that works when asking for 'simpler' does not
- The 26 words that mean your AI is stalling instead of thinking
- The one thing you must never simplify, and why it matters most
Get It Installed First
The whole thing is a free Claude Code skill. It is one command from inside any project, and then you have a /midwit command that runs the pass below on any question or plan you give it.
npx skills add coopersimson96/midwitRepo: github.com/coopersimson96/midwit
Once installed, type /midwit before a scoping question, or say "midwit this" when an answer comes back over-built. Everything below is what it actually does, so you can run it by hand too.
The Problem: The Middle Of The Curve
Picture a bell curve with three people on it. The one on the left is too dumb to overthink the problem, so he says just use a spreadsheet. The monk on the right went through every option, weighed all of it, and says the exact same thing. The one in the middle knows just enough to talk himself out of the right answer. Every AI answer you have ever received landed in that middle position.
It is not stupidity. Detail reads as rigour, hedging reads as care, and structure reads as competence, so a four phase plan looks more considered than a one line answer. The middle of the curve is where an answer goes when nothing is pushing back on it. That is why you ask how to track something and receive a database, an API layer, a dashboard and a nightly job, when you needed five columns in a spreadsheet.
Why Asking For Simpler Fails
Simple is a taste word. Tell a model to be concise and you get three tight paragraphs, then by round three it is back to nine steps. Taste words have no edge, so there is nothing to enforce and nothing to check, and every instruction built on one gets renegotiated. The fix has to be mechanical.
The Rule: Complexity Pays Rent
Every element of an answer that goes beyond the dumbest version that would work must name the exact thing that breaks without it. A valid justification has three parts: a subject (the thing that breaks), a trigger (the condition under which it breaks), and a timescale (now, or a real date). The sentence shape is: without X, when Y happens, Z breaks. If you cannot write that sentence, the element was never load bearing, and it gets cut.
Valid versus invalid, side by side Valid: 'Without the uniqueness check, when two live posts share a word, everyone who types it gets the wrong thing.' It names a failure that has actually happened. Invalid: 'It is more robust.' Robust against what? The test is falsifiability. If there is no future where you look back and say that was unnecessary, you wrote a feeling, not a reason.
The 26 Words That Mean It Is Stalling
Certain phrases exist to let a claim avoid being checked. When one of them is the reason given for a piece of complexity, that piece dies without further debate. They are not bad words, they are unfalsifiable ones, and unfalsifiable is the native language of the middle of the curve.
Paste this into your rules file
Auto-cut vocabulary. A justification built on any of these is void:
scalable future-proof enterprise-grade
best practice industry standard battle-tested
robust flexible extensible
maintainable clean architecture separation of concerns
in case down the line as you grow
you might want to you could consider it depends
for completeness worth noting comprehensive
holistic end-to-end (as a virtue)The Pass, In Seven Steps
- State the objective in one sentence. What do you physically hold when this is done? If the sentence will not close, the request is unclear and saying so is the answer.
- Write the dumbest version that would work. It must be an action. Manual counts, ugly counts, one file counts.
- Write the full version you were about to give, honestly and at full size. This one never ships.
- Diff them. Every element the full version has and the dumb one does not becomes a ledger line.
- Charge each line. Write the justification. Most will not survive, and that is the pass working.
- Rebuild from the dumb version, adding back only what paid rent. Rebuild rather than edit, because an edited long answer keeps the long answer's shape.
- Ship it, then list what was cut so anything can come back in three words.
Why rebuild instead of edit An edited long answer keeps its phases, its ordering and its hedging cadence. The shape is most of the bloat, and the shape survives editing. Starting again from the dumb version is the only way to lose it.
The Prompt
Paste this before any scoping question
Before you answer, run this pass and show me only steps 1, 6 and 7.
1. State my objective in one sentence: what do I physically hold when this is done? If it will not fit in one sentence, ask me a question instead of guessing.
2. In scratch, write the dumbest version that would work. It must be an action, not a caveat. Manual counts. Ugly counts. One file counts.
3. In scratch, write the full version you were about to give me, at full size.
4. Diff them. Every element the full version has and the dumb one does not becomes a ledger line.
5. Charge each line: it survives only if you can write "without X, when Y happens, Z breaks" with a real subject, trigger and timescale. Scalable, robust, best practice, maintainable, in case, down the line, you might want to, and it depends are all void as justifications.
6. Give me the dumb version plus only what survived. Rebuild it, do not edit the long one.
7. Then list what you cut, one line each, no justification, so I can ask for any of it back.
Never cut a verification step, anything irreversible, anything about security or money, a fact, or anything I explicitly asked for.A Worked Example
The ask: a way to see which of my posts perform best. The middle answer: schedule an API pull, store it in a database, build a dashboard with filters and charts, normalise for follower growth, add a nightly job, alert when a format underperforms.
- API automation. Manual entry is 40 seconds per post at one post a day. Cut, that is four minutes a week.
- Database. 'Spreadsheets fall over at scale' has no number attached, and 300 rows is not scale. Cut.
- Dashboard. 'Easier to read' is a taste word. A sorted sheet is also easy to read. Cut.
- Follower normalisation. Without it, older posts rank falsely low purely because reach was smaller then, which inverts the ranking. SURVIVES.
- Nightly job and alerting. At one post a day you notice a bad format in two days. Cut.
The answer is a spreadsheet with five columns, the fifth being views divided by follower count at post time. Notice which piece survived. It is the one nobody would have built first, and it survived because it was the only one naming a genuinely wrong outcome. That is the normal result.
The Floor: What Never Gets Cut
This discipline has one catastrophic failure mode. It becomes the justification for skipping a step, and the rationale always sounds good. That step was overhead. That check was ceremony. We were being lean. So some things are permanently exempt, and they do not need to justify themselves.
- Any verification that proves the thing actually works: loading the URL, opening the screenshot, running the real thing.
- Anything irreversible: deletes, overwrites, sends, payments, anything published externally.
- Security, credentials and permissions.
- Factual accuracy. Fewer words, never fewer facts.
- Anything you explicitly asked for. This shapes what gets added on initiative, never what was requested.
The one line version Simplify the answer. Never simplify the verification. Cutting a verification step is not simplicity, it is unverified work with better marketing.
The Other Ditch
The fool and the monk say identical words, and they are not giving the same answer. The monk went through the complexity and came out the other side. The fool never entered it. From outside they are indistinguishable, which is exactly why the full version in step 3 is not optional. An answer that is short because nobody checked is not simplicity, it is a guess with good posture. Do the full thinking, then report the minimum.
Do This Today
Take the last plan an AI gave you. Count the steps. For each one past the first, try to write the sentence: without this, when that happens, this breaks. Whatever you cannot write a sentence for was never holding anything up. That is usually most of it.