Most AI content tools start too late.
They ask for a topic, a brief, a tone of voice, maybe a few examples, then generate a LinkedIn post. That can be useful, but it is also the easy part. If you give a good model a clear brief and enough context, it can write a decent first draft.
The harder question is upstream: where does the good brief come from?
Fresh is my answer to that question. It is a system I am building for customer work, and today it produces roughly 98% of the LinkedIn content I help clients publish. Not because it magically writes better than humans. Because it captures the useful small things that happen inside company work, then turns them into angles a real person can credibly own.
Draft-first content looks productive
Draft-first workflows feel good because they produce an artifact immediately.
You paste a prompt. The model writes a post. You edit the post. The calendar moves forward. For a few weeks, this feels like progress.
Then the weakness starts to show. The inputs are thin: a product update, a vague topic, a recycled founder opinion, a company announcement, a trend everyone is already commenting on. The output is not always bad. It is worse than bad. It is plausible.
Plausible content fills a calendar while weakening the person who publishes it. It sounds like a company trying to be present, not like someone doing the work, noticing the market, and helping their audience see something more clearly.
That is the distinction Fresh is built around. The goal is not to help companies post more corporate content through personal accounts. The goal is to help people become visible for expertise they actually have, using source material from the work they are already doing.
What Fresh captures
The raw material is everywhere.
A sales objection that keeps coming back. A customer sentence from a call. A product decision in Linear. A Notion spec with a tradeoff. A Slack thread where someone explains the market better than the website does. A support question that reveals confusion. A competitor move. A small internal debate that would make a strong public post if it were framed properly.
Most of these signals are too small to publish alone. That is why they disappear.
One customer sentence is not a post. One Linear ticket is not a content strategy. One Slack explanation is not enough to ask a busy founder to write on LinkedIn.
But when they are captured, tagged, grouped, and connected, they become useful. Fresh treats them as editorial atoms: small units of evidence, observation, proof, friction, or perspective that can support a stronger angle later.
This is the same instinct behind Rift, pointed at a different domain. Rift captures useful memory across AI work. Fresh captures useful editorial material across company work. In both cases, the first problem is not generation. It is noticing and preserving the source material.
The product is assembly
Writing is not the hardest part.
Assembly is.
Who should publish this idea? Is it a founder angle, a sales angle, an expert angle, or a corporate account angle? Does the idea need more proof? Does it fit the person's actual voice and credibility? Would their audience care, or is it only useful to the company's agenda?
Fresh tries to answer those questions before drafting. It looks at sources where work already happens, extracts promising signals, groups related atoms, and turns the best clusters into candidate ideas in a Notion content calendar. A human can accept, reject, ignore, request a draft, or ask for a better angle.
That human loop matters. The system should make editorial judgment easier, not pretend judgment disappeared.
The strongest version is multi-profile. In one company, five people may have five different editorial territories. One person should talk about operations. Another should talk about customer pain. Another should talk about product tradeoffs. Another should talk about hiring or management.
The company benefits from all of it, but each person still builds something personal.
Why this creates trust
There are three kinds of trust in the system.
First, audience trust. People can feel when a post comes from actual work. A concrete objection, a precise tradeoff, a surprising customer sentence, or a real lesson from delivery does more than a polished slogan. Specificity carries the authority.
Second, author trust. Many people inside companies dislike personal posting because it feels fake or too close to marketing. Fresh works better when it shows them material that is recognizably theirs: something they said, saw, learned, or helped solve. The system is not asking them to cosplay as a thought leader. It is showing them the public value of their own work.
Third, company trust. This is the commercial layer, but it should not come first. When several people share useful grounded ideas over time, the company becomes easier to believe. The market sees expertise before it sees an offer.
That only works if the content is useful first.
What is already working
The current version connects to places where work happens, extracts candidate material, and turns it into idea briefs.
Notion is the operating surface because customers already understand it. Fresh can create publication ideas, attach source context, suggest an owner, and provide enough proof that someone can decide whether the idea deserves a draft.
I am still improving the unglamorous parts: source connectors, triage, duplicate detection, owner matching, atom grouping, approval loops, and feedback from published posts.
The customer feedback has been clarifying. When Fresh proposes a weak idea, nobody cares that the pipeline was clever. When it proposes an idea someone immediately recognizes as valuable, the product earns trust.
The lesson so far is simple: angle plus proof beats draft-first generation. A rough draft from a strong angle is easy to fix. A polished draft from a weak idea is still weak.
What does not work yet
Owner matching is still harder than it looks. A topic can be strategically useful and still belong to the wrong person.
Weak signals are noisy. A Slack thread may contain one useful sentence and a lot of context that should never become public. The system needs to extract the signal without laundering private mess into public content.
Proof is uneven. Some ideas come with customer language, numbers, screenshots, or repeated objections. Others come with vibes. Fresh has to push weak ideas back into "needs proof" instead of turning them into confident drafts.
Voice is also not solved by a tone prompt. A person's voice includes what they care about, what they refuse to say, what they have permission to say, and what they have actually earned the right to explain. That takes more than examples.
What I would copy first
If you want to rebuild the useful part without Fresh, do not start by asking an LLM to write posts.
Start with one database for source signals. Capture small pieces from sales calls, customer feedback, Slack, specs, product decisions, market observations, and delivery work. For every signal, write four fields: what happened, why it matters, who could credibly say it, and what proof supports it.
Then group signals weekly. Do not ask "what should we post?" Ask better questions: which signals repeat, which reveal a belief, which would help the audience, which fit a person's actual expertise, which are too corporate, and which need more proof before they deserve a draft.
Only then draft.
That is the design of Fresh. Drafting is downstream. The useful work is turning company work into trustworthy editorial material before the writing starts.