Say It Plainly
Five skills that make an AI tool review or rewrite your writing in plain words, and a private checker that runs in your browser and sends nothing anywhere.
I write for people who are busy. These are the rules I use, written down so an AI tool can apply them to your draft: emails, meeting notes, reports and public posts. A rewrite is told to keep what your text claims exactly where it was and to add no fact you did not give. In my tests it mostly did, but not always. See what I have tested below.
Try the checker: shaunmarsden.github.io/say-it-plainly. Paste some writing and it marks the words worth a second look.
Use a skill
The quickest way needs no install. Open a skill below, click Raw, copy the whole file, paste it into any AI chat, then paste your writing underneath. Say whether you want it reviewed or rewritten.
To install one for an AI coding tool, run for example npx skills add shaunmarsden/say-it-plainly --skill plain-email.
Want something shorter? In my tests a single paragraph kept a draft's claims about as well as a skill did (email and report). On call notes it did better than the first version of the notes skill, and after I tightened that skill the two matched (round two, round three). On longer, messier drafts for the other four skills, the paragraph and the skills lost nothing and invented nothing, where a bare "make this clearer" did both (round five). In that round the report skill added a line on how sure the writer was, which the draft had not said, so check that line if you use it. Paste this above your writing:
Rewrite this in plain words. Put the point first, use short everyday words, cut filler and name who did what. Keep every figure, every hedge and every disagreement exactly as strong as it is in my draft, and add no fact, name, date or owner that I did not give.
That was seventeen drafts and 108 replies, so it is a hint, not a proof. The skills add review mode, rules for each kind of writing, and questions about what they could not tell.
What a skill does
- Review critiques your text and leaves it alone. You get a verdict, the seven changes that would help most (each with the exact words quoted), what already works, and what it could not tell without more from you.
- Rewrite returns the edited text, then what changed, what it kept on purpose, and what it needs from you. It is told to keep your scope, certainty, emphasis and cause, and to keep your weaker claim if a clearer sentence would claim more than your original did.
- Interview (report and post only) asks up to five questions about your reader and what only you know, then drafts from your answers.
Every skill carries all its own rules, because a pasted file cannot open a link. The twelve shared rules are in rules/core.md. Each skill adds the few that belong to its kind of writing.
My own formatting preferences, which I call house formatting (dates like 21st May, numbers one to nine spelled out, no em dashes, straight quotes, full currency figures) are an option in each skill and a checkbox in the checker. They are not rules of good writing, so they are off unless you ask.
What I have tested
- The checker. 31 tests cover every flag, and the cases that must not be flagged, such as a clean email, a year in notes, or a number range. I broke the checker on purpose in 14 ways, for example by removing the overlap handling and by changing the long-sentence limit, and the tests caught all 14. One broken case slipped through the first time, because my test could not have caught it, so I rewrote the test and ran it again. In a browser I checked that pasted markup shows as plain text and creates no elements, that the page's only network request is its own
checker.js, and that the console is clean.
- Whether the skills keep your claims. On 26th October 2026 I ran a pre-registered test: three made-up drafts (an email, call notes, a report), each rewritten twice by each of three methods: a bare request, a one-paragraph instruction and the matching skill. I scored 18 replies without knowing which method each came from. The bare request lost or invented 5 things, the paragraph 1 and the skill 0. The skill and the paragraph were too close to tell apart, and the notes draft separated nothing. The sample is small and I wrote the drafts, the skills and the scoring. The results have the detail, including the judgement calls I made in scoring.
- Whether the notes skill keeps your claims. The first test showed nothing for notes, so on the same day I ran a second, pre-registered round on three messier sets of made-up notes. Every method kept every item, but the notes skill invented 5 things (such as "nothing was decided on this call" and calling a review a "review meeting"), the bare request 2 and the paragraph none. I then tightened the notes skill, in
rules/use-cases/notes.md, and ran a third round on three new sets of notes with four methods: bare, the paragraph, the old skill and the new skill. The old skill invented 4 things and lost 1, the bare request invented 6 and lost 1, and the new skill and the paragraph invented and added nothing. The fix was written for the faults round two found and the new notes were built to tempt the same faults, so this shows the fix works on those faults, not that the notes skill is good. The new skill's rewrites are about 1.7 times the length of the paragraph's. See the round two and round three results.
- The email, report, post and general skills. A fourth pre-registered round, on 26th October 2026, rewrote four made-up drafts (one for each of those skills) by the three methods, 24 replies in all. Every method kept every item, and one bare reply invented a small detail ("our depot" for "the depot"). Nothing separated the skills from the paragraph, and two of my three predictions failed. The drafts were short and easy, so the round found no fault but could not show a difference either. See the .
What I have not tested
- Whether the skills improve real writing. The test above measured only whether claims survived a rewrite. I have not tested whether readers find the results easier to read.
- The published page in a browser. I checked a local copy. The published files are identical to it, and the published checker flags what it should when run in Node, but I have not seen the live page rendered.
- Anyone else's use. Nobody outside this project has tried any of it.
Limits
- The checker looks at words and sentence length. It cannot tell whether you are right or whether your writing is good, and it will flag things that are fine. Each flag is a prompt to look again.
- "Common in machine-written text" is not an AI detector. Plenty of people write like that too.
- Reading ease is an estimate from word and sentence length.
- The passive-voice check is a rough pattern, so it misses some passives and flags some adjectives.
- English only. The skills are 900 to 1,250 words each, because each one repeats the shared rules.
Build and check
The skill files are generated. Edit the files in rules/, then run:
python3 scripts/build_skills.py
python3 scripts/check_repo.py
python3 -m unittest tests/test_build.py
node --test tests/checker.test.js
The same four commands run on every push.
Where this came from
The rules owe most to George Orwell's six rules in "Politics and the English Language" (1946). The idea of a writing skill with a private checker came from Addy Osmani's Clarity. The rules, skills and code here are my own.
MIT licence.
