Four tools · one refusal
A recommendation with no source, a rewrite that still sounds like a bot, a deck nobody can diff, a diagram that breaks on the next render: the usual fix is asking the model to try harder. Clearfelt scores, rewrites, renders and reviews in code instead, so every result is reproducible on your machine and every claim points at where it came from.
npx skills add andrei-ionut-nita/clearfelt-writingFour independent tools, one name: install just the one you need. Actively maintained, MIT licensed.
Our platform offers a seamless, robust solution that delivers pivotal insights. Studies show that businesses leveraging our tools see paramount improvements. In conclusion, this is a game-changer for your organization.
Our platform gives you insights you can actually act on, without the setup headache most tools come with. Teams using it tend to see real gains fast. If your org is stuck deciding, this is probably the push you need.
They share no code. They share a refusal: no LLM grading its own output, no score you can massage, no recommendation without a source you can open.
Scores and rewrites AI-sounding writing against a rule dictionary where every rule cites real research. Runs offline, zero dependencies.
Human score · before
22
After rewrite
100
Static SVG at build time, colors aliased to your own CSS tokens, and a linter built on 12 Cardinal Rules written after 95 real diagrams shipped wrong.
architecture.svg · rendered once, at build time
cardinal-rules-lint architecture.mmd ✓ color semantics ✓ one accent node ✓ nodes per row ≤ 4 ✓ diamond sizing new violations: 0 · baseline-diffed
Author decks as plain .astro files on a 1920×1080 canvas, edit them in a live browser inspector, present from the same app. Exports PDF and PPTX.
<section class="slide bar-compare"> <h2>Time to a reviewable deck</h2> <li data-bar="45">Export, upload, comment</li> <li data-bar="2" data-accent>git push</li> </section>
Time to a reviewable deck
An entity and an objective in, a change plan out. Every recommendation cites a finding, which cites evidence, which cites the page it was read off.
trace R-0002 · recommendation to source
01
No model judges the output. Scores and priorities are arithmetic, printed alongside the result, so nothing claims P0 without earning it.
02
Research that was blocked or fruitless is logged as a gap. "We could not find out" never renders the same way as "we checked".
03
Rules, config, voice, decks and diagrams are plain Markdown, .astro and .mmd files. Hand-editable, diffable, reviewable in a pull request.
04
Opinionated defaults run on an empty directory. Voice profiles, palettes and themes are depth you add later, never a prerequisite.
Each tool installs on its own. Nothing depends on the others.
New here? Start with clearfelt-writing, it works on any file, no setup.
clearfelt-writing
npx skills add andrei-ionut-nita/clearfelt-writing
clearfelt-diagram
npx clearfelt-diagram-init init my-docs
clearfelt-slide
npx clearfelt-slide init my-deck
clearfelt-review
node packages/core/dist/cli.js init acme
MIT licensed. Each repo carries its own README, CHANGELOG and roadmap of what stays deliberately unbuilt.