We build a platform that simulates, grades, and elevates human conversation. When your entire product is human communication, you don't get to be vague about where the machines do your talking for you. So we won't be.
We embrace the bleeding edge. We also respect craft enough to know exactly when the bots need to clock out. This document is the map of that line — public, specific, and updated whenever the tooling (or the line) moves.
The one rule
Everything below is downstream of a single principle:
If a task is mechanical, verifiable, and historically tedious, a model can draft it and a human will check it. If a task requires accountability, taste, or a signature that means something — a legal filing, a payment, a published opinion, a design someone is supposed to feel — a human does it. Confidence is not correctness, and "the model was sure" has never held up in front of a judge, a customer, or us.2
Read this twiceThe Firewall: our tooling is not your data
Read this part twice, because it's the one that actually matters.
This policy describes how we build Intentioned.tech. It does not describe anything that happens when you use Intentioned.tech.
The frontier models mentioned here live in our development pipeline — on our machines, in our editors, during our workflow. They are not wired into the running app.
Intentioned.tech is local-first and privacy-first by design.3 Your conversations and voice data are yours. Our choice to use a cloud model to write a favicon does not mean your practice sessions take a field trip to somebody's inference cluster. Dev-time and run-time are different universes, and we keep a wall between them on purpose.
🟢 Green lightWhere we let the machines cook
We automate busywork so humans can spend their finite hours on architecture, correctness, and taste.
Frontier vibe coding. We use top-tier, state-of-the-art models to accelerate development. Hand-writing boilerplate is a relic; we'd rather spend the cycles on system design and high-level logic.
HTML, markup & scaffolding. The models handle structural markup and config. Manually fighting to center a div is a waste of a human lifespan.4
Vector assets — SVG logos & favicons. We use AI to write the math behind clean, scalable vectors. An SVG either renders correctly or it doesn't; there's a right answer, and a human confirms we hit it.5
Deep systems integration. We lean on AI to scaffold complex backend work — Transformers architectures, CUDA/ROCm integration, and the hardware-acceleration plumbing nobody enjoys wiring by hand. AI drafts the pipes; humans own whether they hold water.
Blog & copy assistance. AI is our sounding board and grammar checker: outline, brainstorm, tighten, catch the typo. The thinking is ours. The model just holds the mirror.
Research & analysis (verified). First-pass research, source summarization, and draft business or competitive analysis. Hard rule: nothing factual reaches you until a human has checked it against the source. A confident paragraph is not a citation.
🛑 Hard noWhere the bots clock out
If it compromises the product's integrity, its accountability, or its soul, it's banned. No asterisks except the one we explicitly wrote down.
Cheap-model vibe coding. If we're vibe coding, it's the heavy hitters — full stop. Budget-tier, non-frontier models that confidently hallucinate garbage into the repo are forbidden.6
Generative raster art. Vector math is fine. Generative raster art is not. Real artwork — commissioned, human, intentional — is on the roadmap for reward stimulation, and it's worth doing right.
Full AI-written blog posts. AI does not write our articles; we write our own thoughts. The single exception is a code-red, time-critical update where communication speed is the only metric that matters — and even those ship with a visible "AI-drafted, human-reviewed" tag.
Slop & plagiarism. No soulless content-farm filler, no lifting anyone else's work. If an output reads like it was extruded rather than written, or it borrowed without credit, it goes straight to /dev/null.7
Legal duties & payment processing. AI does not touch the courts,8 and it does not touch the cash. Legal obligations and payment flows run on human accountability, because neither PCI compliance nor a judge's patience accepts "temperature was set to 0.7" as a defense.
Known AI blind spots. If LLMs are historically bad at it — nuanced subjective calls, the stuff that needs lived intuition, anything a model can't verify and neither can we — we don't force it. Human judgment and real testing win those.
AccountabilityHow we keep ourselves honest
- Every public-facing artifact gets human sign-off before it ships. No exceptions.
- Substantially AI-drafted public content is labeled as such. You should never have to guess.
- This policy is versioned and dated. When the tooling changes or we change our minds, the change shows up at the top — where you can actually find it.
Think we've drawn a line in the wrong place? Tell us. We'd rather be corrected than comfortable.
Footnotes
- "Living document" is not a euphemism for "we'll forget to update it." It has a date at the top. Check it. ↩
- A language model's certainty is a property of its output distribution, not of reality. These two things are unrelated more often than is comfortable. ↩
- "Local-first" means the app is built to run on your hardware without a mandatory round-trip to someone else's servers. "Privacy-first" means that's a design constraint, not a marketing adjective. ↩
- Somewhere, right now, a competent engineer is losing a fistfight with
margin: auto. This is the one thing we are all glad to delegate. ↩ - An SVG is deterministic geometry — coordinates and curves that resolve the same way every time. A raster art model is guessing pixels. One is reproducible; the other is a slot machine with better branding. ↩
- "Frontier" is a moving target — today's state-of-the-art is next year's baseline. The point isn't the label; it's the tier. A cheap model is not a discount frontier model. It's a different animal that will invent an API which has never existed and hand it to you with a straight face. ↩
- We are not going to formally define "slop." Like a certain Supreme Court justice on a certain other category of content, we know it when we see it. ↩
- Yes, including this one. Legal and policy documents — this page included — are written with AI and then checked, line by line, by a human who is personally on the hook for what they say. The app's legal and compliance systems work the other way around: coded by humans first, then handed to a frontier model to pressure-test for compliance. Our standing order for both is max effort and the newest, best model available — and no, we will not be defining "newest and best," because it moves faster than this document does. And when the compliance is genuinely hairy — the multi-jurisdiction, cross-referenced kind where a tired human misses a clause and never knows it — we deliberately chug it through a frontier vibe-coding pass rather than skim it by eye and hope; more machine, in that case, is the careful option, not the lazy one. What AI never gets is the duty itself. Drafting the language and checking the code is assistance. Putting a name behind it is accountability, and that stays human. ↩