I’ve been using AI for content output for over three years, and this is what it does every time it writes copy. For example: you ask for a campaign recap, an email draft, or a one-pager, and what comes back is the thing you asked for engulfed by a sea of sloppy content you didn’t ask for — irrelevant considerations, redundant points, needless details, bad grammar, and content direction — that you now have to whittle down and refine so much that by the time you’re done, you could have written two blog posts yourself.
I wrote about this in Welcome to A.I. Basecamp: working with AI sent my mental load up, not down, because “I was also managing a slew of A.I. agents and their prolific, exceptionally verbose output.” And verbose doesn’t begin to cover it. The output isn’t just long — it’s long *and* sloppy:
- Run-on sentences next to sentence fragments
- Headings that say nothing
- Two bullets making one point
- Jargon where plain words would do
- Confident claims that fall apart the moment you check them
”Make this shorter” and “simplify” are two different requests
When the overload hits, my instinct is to type two words back at the AI: *shorter*, and *simpler*. It reliably delivers only the first.
I ask for shorter, and the word count drops — but the meaning drops with it, because the words AI cuts first are the small ones: the *because*, the *though*, the *which means*. They carry no facts, so they look like padding. Except those are the words doing the arguing. With them gone, what’s left is a stack of true statements sitting next to each other like strangers at a bus stop, and I’m left reassembling the logic myself — the exact job I was trying to hand off in the first place.
And sometimes shortening doesn’t just flatten the point — it inverts it. An early draft of this post made a claim in its intro that its own argument, two sections later, flatly contradicted. Both sentences were clean, confident, and beautifully scannable, which is exactly what makes this failure so hard to catch: a tidy sentence reads as authoritative whether or not anyone checked it. Shorter and wrong is worse than long and right, because nobody proofreads the version that sounds sure of itself.
Knowing how AI output fails is the skill
Over three years, every time I caught one of these failures I wrote down the fix — what the failure looked like, how to spot it, and what to do instead. That list turned out to be the real skill — not “using AI,” but knowing every way its output goes bad and having a fix ready for each.
`make-it-make-sense` is that list, packaged as a skill I run as a command: `/make-it-make-sense`. Point it at any draft output, and it runs the full editing pass.
What the skill actually does
Here’s the core of it, straight from the skill:
**The goal is clarity — understandable on the first pass.** And because AI output is almost always too detailed, too dense, and pitched too deep, simplifying is the default move — though not the only one. Sometimes clarity means restructuring, reframing, or adding the missing signpost.
**Core skill principles:**
- **Lead with the answer** — the main takeaway goes first, so the reader gets the point in seconds instead of after four paragraphs of buildup
- **Cut to what matters** — keep the few points that carry the message; everything else gets dropped or moved to an appendix
- **Plain words** — jargon and acronyms get swapped for what they actually mean; if a term has to stay, it gets defined once
- **Don’t overshoot into bare** — every point that survives still needs a real anchor: a number, a name, an example
**Headings are either a claim or a label — never the limp in-between** that gestures at cleverness without saying anything. A heading either makes a point that the following section pays off, or it’s a plain label like “Glossary.” The test, straight from the skill: read only the headings, in order. They should tell the story on their own.
**A redundancy pass across the whole piece.** Every section gets checked against every other with one question: does the reader learn something new here, or did they already get this? The chart and the table showing the same numbers, the intro that restates the summary — merged or cut.
**And the constraint on all of it: the shorter version has to stay true.** The skill names the specific ways simplifying lies — quantifiers that creep (“three segments” becomes “every segment”), verbs that collapse (a step that *drafts* an email becomes one that *sends* it), advice that flips (“review carefully” compressed into “trust the output”) — and it flags anything it can’t verify instead of shipping it clean and confident.
Grab it, break it, tell me what you find
`make-it-make-sense` isn’t trying to be more than it is: a list of editing moves that saves me from making the same corrections by hand every day. It lives in jonstermash-skills, a public repo, so anyone can use it. If it fixes something for you — or misses something — I’d genuinely love the feedback. More skills are on the way.
# # #
P.S. I asked Claude to help me write this post.
Yep, this one. A post about a skill I built specifically to stop AI from overproducing — which, of all the assignments you could hand it, you’d think would be the one it approached with a little humility.
I gave every effort a D-.
Most of the drafting happened on Opus 4.8; I eventually had to switch to Fable 5 to get what now seems like a decent output.
Here’s a sampling from the first draft:
- A metaphor about a topographical survey that it kept propping up long after it stopped working
- Sentences. Chopped. Into. Fragments. For rhythm.
- A section head called “Why a file and not a prompt,” which means nothing to anyone alive
- An entire paragraph dedicated to explaining that a text file is a file with text in it
- Nearly twice the word count of the version you’re reading now
This just underscores the need for this skill. I also added functionality that you can use to give you suggestions for your account output styles to further improve your AI results.