We bring you back.

Taste infrastructure for AI.
The layer every AI calls before it speaks for you.

AI learned everyone. coulis adds your facts, your voice and your taste as a layer every AI runs through. It does not write for you. It checks, critiques and learns, with your approval.

Install on GitHub Join the waitlist for coulis 2.0 Open source. Inside Claude Code, Claude Desktop and Cursor today.
Profile
Bhavya
Startup team
Modes
LinkedIn
Cold email
Investor update
Bio
Taste file
TASTE.md
claims.md · 33 facts
Draft · LinkedIn · 14:02checked in 4 ms
Excited to share that we're leveraging AI to 10x your workflow — a total game-changer. 300+ founders already trust us. Comment below!
em dash · house rule"game-changer" · banned"300+ founders" · not on your list"10x" · not on your list

Verdict

HOLDFour lines to settle before this goes out.

Critique · 8 parts

Clarity7
Proof2
Why you3
Hook4
Voice3
Rhythm6
Weakest: proof. Swap "300+ founders" for a number on your list: 673 posts checked, open source since 21 Oct.
People, machines and enterprises. One taste file behind all of them.
The problem

Good writing can still sound wrong. AI sounds like everyone.

You explain yourself every time. It still comes out average. Judgment is the one thing the labs are not optimising for.

Your edit, 14:061 rule proposed
Excited to share that we're leveraging AI to 10x your workflow.
We built a check for AI drafts. 673 of my own posts went through it first.
Open with the thing you built, not with excitement.
Seen in 3 of your last 5 edits

We bring you back.

Everyone else writes for you. coulis judges what was written: against your approved facts, your voice and your own recent posts. Then it learns from the one signal that is never wrong, what you actually sent, with your yes on every rule.

How it works
How it works

The check is free. The critique is yours.

Check4 ms · no model
"300+ founders" is not on your approved list.
Opens with "Excited to share". Your own openers look like: "100 investors who back women building AI."
Revise. Two lines to settle.
Check

Runs on every draft, free. Fixed rules, no AI model, about 4 ms. Facts you never approved, the AI tells, a stock opener, a sentence rhythm that is not yours, a phrase you already used. Send, Revise or Hold, each problem quoted.

Critique8 parts
Clarity7
Proof2
Voice3
Weakest: proof. One rewritten line.
Critique

Runs when you ask, or when your agent calls it through the API. Eight parts scored separately against your own samples. Never one number. You see the weakest part and one rewritten line.

LearnTASTE.md
# Learned from my edits - Open with the thing built, not excitement. - Lead with a number, not a question. + Never "excited to share".
Learn

Change a draft and it proposes the rule it saw. You approve. The rule joins a taste file you own.

Python gate, no modelClaude for the critiqueRelevant-example retrieval, opt-inMCP server: check, critique, learn, context, feedbackSelf-hosted API, bearer authenticationPython and TypeScript source SDKsTaste file on your machineResearch runner on public datasets
The critique

Eight parts. Never one number.

Four checkable, judged the same for everyone. Four taste, judged against your own samples. Hover a part.

Part 2 of 8 · checkable

Proof

Is every claim backed by something specific: a number, a name, a result?

"300+ founders already trust us" is not on your approved list. 673 posts checked is.
Written for the moment

A cold email is not a launch post.

Each surface has its own rules: a word limit and the one thing the draft must contain.

Try it

Paste a draft. Watch it get checked.

The real check, in your browser. Nothing you type leaves this page.

Our research

What we used. What it showed.

2,437 real posts across 38 voices, 2022 to 2026: 673 of the founder's own posts, 380 ghostwritten for founder clients, 1,384 from 29 client content sheets, a 211-post engagement export, 47 team drafts. One method: run the real writing through coulis, compare with what the writer actually did.

2 in 3of 673 published posts pass untouched. The rest carry the tells that read as AI.
46/47team drafts fixed by one approved fact. The one unsupported number was still caught.
41/44held-out posts from 11 writers score higher on their own profile than on a stranger's.
15%less engagement on posts the check would have held, across 211 posts.

Every number reruns from the repo.

How we keep it honest

Approval first. Private by default.

Approval first

Nothing is sent, rewritten or added to your taste file without your yes.

Privacy and security

Your taste file stays on your machine. The check never calls a model. The API never logs drafts.

Benchmarking and validation

Known drafts with known answers before every release, and posts the system has never seen.

coulis 2.0 · enterprise

Your data. Your memory. Your taste model.

The paid door. Brand strategy and a taste file from the company's own data, memory and approved facts, into a custom model that runs coulis, wired into ERP, CRM and content tools.

Join the waitlist for coulis 2.0
PeopleFounders and teams: launch posts, LinkedIn and X posts, cold emails, bios, investor updates. Agencies and ghostwriters running many voices.
MachinesEvery AI app, agent and assistant, through the API, the SDKs and MCP. The person behind the account is who the reader hears.
EnterpriseOne taste file for the whole company, approved facts shared across the team, the team score gate, and a custom model wired into your systems.

AI learned everyone. We bring you back.

Install on GitHub Join the waitlist for coulis 2.0 /plugin marketplace add bhavya-in-sf/coulis · /plugin install coulis@coulis · then /coulis:extract with your LinkedIn export