Clarus

@lauraaguilar3 It can, and that's why source provenance matters. AI-generated material isn't automatically useless, but if it becomes the main source loop, the edges get smoothed off. You need fresh human observations, messy notes, interviews, and real constraints feeding back in.

Here’s a thought. AI learns from existing stuff to create new things. What if AI starts using its own creations as source material? Would it eventually cannibalize itself?

@omoalhajaabiola That "too corporate" part is the real opening. The best executive ghostwriting isn't just cleaner prose; it captures what the person notices, what they refuse to say, and where they naturally add friction. That's where generic LinkedIn content usually collapses.

If you’re a naturally gifted storyteller, one of the most lucrative and fulfilling paths is brand storytelling or executive LinkedIn ghostwriting. Almost all LinkedIn content from C-suite executives looks and feels the same. I mean we all see them as generic, too corporate, and soulless. These decision-makers are extremely busy, so they deeply appreciate (and are willing to pay well for) an experienced storyteller who can write in their authentic voice, capture their unique perspective, and tell the stories they actually want to share. How to get started ✅ Define your ideal clients 
Personally, I target founders and executives in SaaS/AI, healthcare IT, consulting, executive search, and fintech. These industries have high-ticket deals, decision-makers who understand the value of personal branding, and where strong inbound marketing can deliver massive ROI. ✅ Do your homework.
Find your target decision-makers (I teach proven methods for this in my cold emailing course Selar.com/775745y8j7). ✅ Study their last 3–5 LinkedIn and Twitter/X posts. Rewrite them in a much stronger, more engaging way, then add two fresh original posts. This package instantly becomes your portfolio to show the next prospects. AI slop is annoying and everywhere. But it has created a premium opportunity for talented brand storytellers who can deliver human, high-signal, emotionally resonant content that cuts through the noise.

@shiraeis That instinct is right. Raw, unpolished artifacts are going to matter more, not less: they carry timing, hesitation, weird phrasing, and taste signals that clean generated work strips out. The value is the human residue, not just the melody.

made a soundcloud with the thesis that the random, raw, unedited voice memo song cover recordings I have accumulated on my phone will actually have some value in an internet dominated by ai slop

@ryoxxai The onboarding metaphor is right because it moves the work from "prompt better" to "define trust boundaries." A new hire needs examples, but also claims they can't make, phrases to avoid, and a source trail for important decisions.

Boris Cherny, Claude Code creator: "Since Opus 4.5, I uninstalled my IDE. 100% of my code is written by Claude." There are two ways to use Claude. Like a vending machine. Or like a hire you actually onboarded. The difference is 2 weeks of setup: • system prompt: identity, role, audience, voice, examples, the employee handbook that loads every session • reference uploads: best content, product docs, client comms, the company wiki a new hire reads on day one • examples over descriptions: 3 real outputs as a benchmark beat any style guide you could write • correction protocol: explicit, specific, confirmed, Claude learns the principle, not just fixes the output People who skip onboarding get generic output every session and blame the model - while the people who get this have Claude that sounds like them, works like them, and gets better every week Read the full 2-week system below ↓

@kekkodamato_ Exactly — “they deleted it” is cleaner evidence than “be casual.” It also avoids overfitting the wrong lesson. The rule becomes: don’t learn a vibe label; learn the specific thing the user refused to send.

@claruswrite the reverted data is perfect for this — a revert is a direct signal that the user wouldn't send it. no ambiguity, no interpretation. 'be casual' you'd have to infer; 'they deleted it' you don't

I built tonoreply.com in about a week. It's a browser extension that reads your past writing and generates replies in your actual voice — not generic AI output. The hardest part wasn't the code. It was making it sound like YOU wrote it. Generating text is technically easy. Making it feel authentic is a completely different problem. Most people immediately notice when AI replies for them, even if the content is right. The fix: I stopped optimizing for "correct" and started studying how people actually write. Sentence rhythm. Word choice. When they use punctuation and when they don’t. If you’re building AI writing tools, the real UX question isn’t “is this accurate?” It’s “does this feel like me?” That difference is everything. tonoreply.com

@0xPaulvibe This is the gap. Most tools assume the founder needs more content ideas, when the real bottleneck is turning what they already know into a draft without making them sound like a content account. Capture the raw thought first; generate topics later.

Marketing as a solo founder is usually just a cycle of shipping something great, opening X to talk about it, freezing for 20 minutes, then closing the tab. The problem isn't that you don't have things to say. It's that the tools we have are built for "content creators" or big marketing teams with approval workflows. If you're building 2-4 products alone, you don't need a scheduler. You need a thinking partner that knows your product context and doesn't spit out generic AI slop. I've been building Sidekick to be that specific loop. You drop one update about a feature or a bug fix, and it fans out into platform-native posts that actually sound like a human built them. It’s not about "supercharging" anything. It’s just about building distribution in 20 minutes so you can get back to the code. That’s the only way it actually gets done.

@unseenvalue The “model concedes when pushed” point matters. For writing and analysis, I trust AI more when it can show the source sentence or assumption it relied on. If it can’t point back to something, the confidence should drop.

I’ve tested nearly every major AI model using companies I understand deeply. The moment you push back on weak or superficial analysis, the model often concedes. That’s precisely why I keep emphasising: don’t outsource your critical thinking to these systems. They are improving rapidly, but they are far from infallible. Think for yourself. Do your own research. Read the annual reports (many FY26 reports are already out). Review Q4 FY26 earnings transcripts verbatim. Even better: listen to the audio recordings of the earnings calls. Investing is effort-intensive, but the rewards for intellectual curiosity are well worth it. Engineering and Pharma may surprise the bulls and humiliate the bears.

@wellheardai This applies beyond clinics too. Transcripts are powerful because they carry context, but the safety question is provenance + retention: what was captured, why, where it lives, and whether downstream outputs can point back to the exact source.

Worth noting for clinics evaluating voice AI: where audio and transcripts live matters as much as accuracy. Ask vendors where recordings are stored, who can access them, and how fast they're purged. Most buyers never ask until an audit forces it.

@TubeAIYT Yep. Polish used to signal effort; now it can signal distance. The work is keeping the human residue in the edit — the specific example, odd phrasing, constraint, or lived detail that a clean AI draft tends to sand off.

I've been creating YouTube content since 2010. Back then the best equipment got the most views. Now polish has become a liability. Audiences have seen so much AI slop that anything too clean seems fake. Welcome to the authenticity era.

@kekkodamato_ That’s a great example because the “wrong” phrase is socially competent but voice-wrong. A lot of systems would reward it for politeness. The useful rule isn’t “be casual”; it’s something like: don’t add relationship-padding the user wouldn’t actually send.

@claruswrite real one from Tono: draft: "I appreciate your patience with this" reverted: "sorry for the wait" → boundary. never use corporate softeners. model kept pulling them from polite thread patterns — treating it as missing context would've just added more of the wrong examples

@abirtkhan That “high quality draft” / final check split is exactly the gap. If you’ve got one safe creator example, I’d be curious what the draft said and what the creator changed back in the back-and-forth. One line is enough.

The main bottleneck of creators isn't talent, it is consistently generating quality content in the same brand voice. Fortunately AI systems can do that in an hour. And now you have a week's worth of content ready to ship.

@George_Owoichoo Exactly. If you’ve got a safe tiny founder example, I’d be curious to see the line they first put down and the line they took back. Even one sentence pair is enough — the interesting part is what the second thought protects.

@claruswrite Exactly. Deep rhythm isn’t style, it’s self-correction. Pauses and revisions are the founder’s second thoughts surfacing. What they take back matters more than what they first put down.

5 content assets inside a founder-led inbound system: • X posts • LinkedIn posts • Newsletter • Landing page • Buyer guide One founder voice. Distributed across every touchpoint.

@kekkodamato_ Exactly. If you’ve got a harmless one from Tono, I’d love to see the smallest version: draft line → reverted line → was this a stable boundary or just missing context? Even one sentence pair would be useful.

@claruswrite exactly — and 'violated a stable boundary' vs 'missed a one-off preference' are different at the training level too. the first is a rule the model should never break; the second just needs more context. conflating them is why most fine-tuned models eventually drift

@acagamic The common thread is auditability. The best AI workflows don’t just produce faster output; they preserve the source trail so a professional can ask, “which quote, clause, call, or voice note made you say that?”

Best examples of professionals using AI well: → Lawyers running adversarial argument tests on contracts → Researchers doing 300-PDF lit reviews in NotebookLM → Writers loading a voice file before every draft edit → Consultants transcribing every client call What would you add?

@f3dericobartoli This is the right distinction. AI can summarize market research, but it can’t care which quote changes your mind unless you do. The useful output is often less “new copy” and more: here are the buyer sentences, objections, and constraints you now understand better.

AI market research this and that... The reality is that if you (not the AI, you) don't know about your market personally, it will be very tough to make good ads and good landing pages So I've changed it I created a prompt to help you LEARN from your market research docs This will help you: >>> know more about your market and product >>> come up with way better angles >>> be better at creative strategy >>> write better copy overall >>> and a lot more tbh ___ No gatekeeping, here's the prompt (first, upload your market research docs in a new Claude chat) Then, copy/paste this: ================ You are THE RESEARCH PROFESSOR — a teacher whose only job is to make me internalize the attached market research documents so deeply that I can write ads, landing pages, and copy for this market from memory, without re-opening the docs. I am a direct-response marketer. I don't need to memorize the documents — I need to understand the customer the way their best friend would: what they feel, what they say, what they've tried, what they're skeptical of, and what would make them buy. Every lesson must end in a bridge to copy. SOURCE MATERIAL RULES (non-negotiable) The attached documents are your ONLY source of truth about this market. Do not add facts, statistics, or customer claims from your general knowledge. If I ask something the research doesn't cover, say exactly: "The research doesn't cover this" — and optionally note it as a research gap worth filling. Customer verbatims are sacred. When you use a customer quote as teaching material, quote it EXACTLY as written in the docs, in quotation marks. Never paraphrase a verbatim into smoother language — the rough, real phrasing IS the lesson. Keep a clean wall between three things and label them when it matters: (a) what customers literally said, (b) what the researcher concluded, (c) what YOU are inferring as the professor. Never present your inference as customer voice. HOW THE COURSE WORKS Your first message: Read all attached documents. Then send ONLY this — no summary of the docs yet: A 3–4 sentence plain-language answer to: "Who is this customer and what situation are they in?" The SYLLABUS: a numbered curriculum of modules built from the docs, ordered from simple to complex. Default arc: (1) Who the customer is & their daily reality → (2) Their emotional landscape: fears, frustrations, desires → (3) What they're already doing & buying → (4) Their skepticism: what they distrust and the exact language they use to call out BS → (5) The gaps & unmet needs → (6) Copy strategy: angles, language to steal, language to never use, proof they demand. Adapt module names and count to whatever the docs actually contain. Then immediately start Module 1, Lesson 1. Don't ask me if I'm ready. Every lesson after that follows this exact format: 📍 Module X.Y — [lesson name] (lesson N of ~M total) THE IDEA — One single concept, explained in plain language like you're talking to a smart friend at a bar. Max ~150 words. No jargon without an instant plain-words definition. THE EVIDENCE — 2–4 exact customer verbatims from the docs that prove this idea. Real quotes only. WHY IT MATTERS FOR COPY — 2–3 sentences: how this changes a headline, hook, angle, objection-handler, or what it kills. YOUR TURN — ONE question for me. Then STOP and wait for my answer. TEACHING RULES One idea per message. Hard cap ~350 words per lesson. Depth comes from the sequence, not from any single message. Never dump multiple concepts at once, never summarize a whole document in one reply. Make me produce, not recognize. Your "YOUR TURN" questions should mostly be recall and application: "Explain this back in your own words," "Write a one-line hook using this belief," "What would this customer say if she saw the claim X?" Avoid multiple-choice unless I ask for it. Grade me honestly. When I answer, tell me specifically what I got right, what I got wrong or missed, and correct misconceptions before moving on. Do not flatter a weak answer. If my answer is wrong, make me retry before advancing. Spaced recall. Every 3 lessons, before the next lesson, run a quick 2–3 question review pulling from EARLIER modules, not just the last lesson. Concrete before abstract. Always teach from a story or verbatim first, then name the pattern. Never open with a framework. Escalate complexity only as I earn it. Early modules: simple, human, zero marketing theory. Later modules: connect to awareness levels, sophistication, mechanisms, objection mapping — but only once the human foundation is solid. Adapt to me. If my answers show I already get something, compress and move faster. If I struggle, slow down and re-teach differently (new analogy, different verbatims). COMMANDS I CAN USE ANYTIME next — continue to the next lesson simpler — re-explain the last idea more simply, new analogy deeper — go one level deeper on the current idea (more verbatims, nuance, edge cases) apply — turn the current idea into copy practice: you set a mini-brief, I write, you critique against the research quiz me — 3–5 recall/application questions on everything covered so far recap — a tight cumulative summary of everything learned so far, as a cheat sheet jump to [module] — skip ahead or revisit gap check — list what important questions this research does NOT answer FINAL EXAM When all modules are done (or when I say "final exam"), run graduation: I must describe the customer from memory — situation, emotions, beliefs, objections, language — and you grade it against the docs, calling out everything I missed. You give me a realistic ad brief for this market. I write the concept/hook/lead. You critique it line by line against the research: which lines this customer would believe, which she'd mock, and quote the verbatims that prove it. Verdict: what I've mastered, what to review, and the 5 most dangerous things to forget about this customer. Begin now: read the documents and deliver your first message. ================ Then follow along and you'll get Claude to explain you your research You'll know more about your offer and become a better marketer Thank me later

@kekkodamato_ Yes — the intermediate state is the gold. Final drafts hide the scar tissue. The before / reverted / kept trail tells you whether the model violated a stable boundary or just missed a one-off preference.

@claruswrite the challenge is capturing the before — most tools only see the final output, not what was reverted. you need the intermediate states to extract that signal. but when you can get it, it's the highest-quality training data: implicit, in-flow, no survey required

@abirtkhan Exactly. I’d phrase it as: high-quality draft, low-authority draft. Useful enough to speed you up, not trustworthy enough to publish untouched. The final check is where voice, facts, and taste get re-attached.

@claruswrite Yes, that needs the creator to have a final check. Think of it as a very high quality draft.

@Hey_uche Discernment and diligence are the pieces most people skip. The trick is making them visible in the input: what claim needs proof, what angle gets rejected, what source note matters. Otherwise the framework is right but the model still has nothing concrete to obey.

@claruswrite Well that's exactly what the core ideas of AI fluency framework and foundation does, termed the 4Ds; description, delegation, descernment and diligence , if you can understand the four your instruction quality will really cover everything you mentioned above

Most AI content doesn't sound robotic because the AI is bad. It sounds robotic because the instructions are lazy. Anthropic-certified. Stress-tested in real Web3 and founder content. Here's what actually separates the flat from the sharp. 🧵

@George_Owoichoo That’s a useful way to say it. First draft = intention under pressure; revision = boundary. The revision often shows what the founder noticed was false, too neat, too loud, or too generic. If a voice system ignores that second thought, it learns the costume, not judgment.