Tutorials
Videos that teach a real AI task
The news tells you what changed. These videos show you how to do something with it. AI Lately adds a few at a time, watches each one end to end, and tests it two ways before it runs here. Can a beginner follow it? Does it name the risks? A creator's audience size decides nothing.
Each video plays here from its maker's own channel, with a link back. The notes under it are AI Lately's, checked against the video's captions.
Get the same answer from a model every time by defining your terms
How to make LLM’s Deterministic like Jev from an AI Researcher
Elin Nguyen tried to sort customers into high, medium and low value with a model. The same customer flipped from high to low overnight with no change in the data. She explains why with a roulette wheel: the model's odds for each next word are fixed, and the spin is not. Her fix is to define every term in the prompt until four vendors' models give the same answer ten times out of ten.GTM AI Academy on YouTube What you can do after it
- Tell the two parts of a model's output apart: the fixed odds it computes for each next word, and the random pick that follows.
- Test one prompt on several models. If the answers differ, the prompt is loose, not the model.
- Define each term until the answers converge. Her tomato prompt had to say what a fruit is, what a vegetable is, and from whose point of view.
- Hold the 10-out-of-10 standard. She calls 90 percent right 'an indicator of 10 percent that's going to be a surprise'.
- Drop the tricks. 'Think step by step' made her car-wash test worse, and a role prompt moved it about 2 percent.
- Watch the small words. Adding 'answer with exactly one word' flipped Opus from the right answer to the wrong one.
The path through it
- 5:00The problem: customers that flip from high to low value with the same data.
- 10:10The roulette wheel: fixed odds, random spin.
- 13:15What she calls substrate engineering: define until the model can only land on one answer.
- 17:20Interpretation drift: 'fruit' against 'the red apple in Snow White'.
- 20:30Jev, and why people want a classifier that answers yes or no.
- 23:20The car-wash test: 17 models tell you to walk to the car wash and leave the car at home.
- 24:30Tomato: vegetable or berry, and why saying 'botanical' was not enough.
- 29:00Her advice for people who prompt every day: test across models and find where the flip-flopping stops.
- 31:40The prompt tricks that barely move the answer.
- 33:30'Answer with exactly one word' flips Opus.
- 39:00Why she calls it meaning engineering: define 'qualified lead' before the model does.
The risks
Named in the video
- Unstable labels in revenue systems. A high-value customer on Monday is a low-value one on Tuesday, and your emails follow the label.
- Trusting a 90 percent score. The other 10 percent is where the surprise lives.
- Instructions that are satisfied and still wreck things. Her example is a model that 'nuked the database' because the wording allowed it.
- Lost context. A model holds one page at a time, so by page ten it may be solving a different problem.
Left out, so AI Lately adds them
- Her evidence is her own. The papers and the 17-model car-wash test are hers, and the video shows no outside check of them. She says much of prompt science is correlation dressed as cause; the same test applies to her work.
- The method is slow. Her tomato spec for 17 models 'was huge'. The video gives no sense of how long a real spec takes or who maintains it.
- Vendor change. A spec tuned to today's models may drift when the vendors update them. The video does not say how often to retest.
AI Lately's test
- Can a beginner follow it?
- Yes. It is a conversation, not a screen share, but every idea comes with an example a beginner can repeat: tomato, car wash, fruit.
- Does it cover the risks?
- Yes. Risk is the whole subject. It names the failure modes and gives a standard to test against.
The title names Jev, a classifier model from TypeSafe. AI Lately's own desks use Jev for small checks under a spending cap. Nguyen says in the video that Jev was down the morning they recorded, so she could not compare it with her method on air.
Set up Claude Code skills and a one-agent-per-task content pipeline
He Built a $500K Business Solo With Claude Code (No Coding)
Mark Fershteyn works in the Claude Code terminal and shows how he builds skills. He starts in plan mode and gives each agent one task. Five or six of them chain from idea to draft to edits to reposts. Then he shows Zealos, the app he built on Claude Code, Vercel and Supabase. It pulls five calendars, every inbox and every message into one screen, and it works out who is visiting his website.GTM AI Academy on YouTube What you can do after it
- Press Shift+Tab in Claude Code to switch into plan mode, and start every build there so the thinking happens before the execution.
- Give each agent one task with a clear pass or fail: one edits, one fetches mail, one triages it.
- Write one skill that runs the agents in order, so a content pipeline becomes a single command.
- Brief a skill with examples the way you would brief an intern. He built one from 300 of a LinkedIn writer's posts; it interviews him, drafts a post in that style and scores five hooks.
- End every skill with three questions: what went well, what did not, and what to change next time, so the skill improves itself.
- Keep every working folder in a GitHub repo as a backup. His clients folder holds each client's emails, Slack threads and LinkedIn messages.
The path through it
- 8:15Claude Code in the terminal against the app: he runs nine small terminal windows at once.
- 9:24Use AI to use AI: talk the problem through in plan mode first.
- 10:26One agent per task.
- 11:31The prompt that builds a set of interview, drafting, editing and repurposing skills.
- 12:20The host's rule: about four rounds of feedback before a skill is good.
- 13:07Treat the agent like an intern: five examples beat one instruction.
- 13:30The skill built from 300 LinkedIn posts.
- 14:08Make every skill ask what to improve at the end.
- 16:00Folders, GitHub and the clients folder.
- 17:08A personal wiki built as a graph, not a hierarchy.
- 18:58Zealos: the app he built on top of all of it.
- 21:58Identifying website visitors and enrolling them in outreach sequences.
- 24:20The revenue claim: a run rate past $500,000 in six months, solo and part time.
The risks
Named in the video
- A one-shot skill is too broad or too thin. The host says it takes about four rounds of output, test and feedback before the results hold.
- Build happiness. He expects a few years in which companies build everything themselves. Then, he says, they will see that putting engineers on an internal CRM cost more than paying for one.
Left out, so AI Lately adds them
- Other people's words. The LinkedIn skill was built by scraping 300 posts by a named writer to imitate his style. The video does not raise LinkedIn's scraping terms or what the writer would think.
- Visitor identification. Zealos works out who is on his website through several data vendors and enrolls them in outreach. Nobody asks whether those visitors agreed to that, or which privacy laws apply.
- The revenue number is his own. The $500,000 is a run rate he states. No figures appear on screen.
- Agents that read every email, Slack thread and client message keep a lot of private data in local folders and a GitHub repo. The video says nothing about who can reach that repo or how it is secured.
AI Lately's test
- Can a beginner follow it?
- Yes. The skills section is step by step: the key to press, the rule to follow, the prompt to type. The Zealos tour that follows is a demo, not a lesson.
- Does it cover the risks?
- No. It covers the craft, iterate and give examples, and nothing on consent, scraping or data security.
Run a sales team's reporting from Claude with connectors and scheduled tasks
200 AI Agents to 10: How a Founder Runs 90% of GTM in Claude
Eli Portnoy shows the Claude setup that runs most of BackEngine's sales work. Connectors give Claude his email, calendar, call recorders, CRM, Notion and Slack. Scheduled tasks email him an outreach pulse at 4 p.m. and list the meetings he has not followed up. Another reviews every 20 prospect calls for what lands and what does not. He once had 200 of those jobs running. Nobody read them. He cut them to about ten, each with an owner and a meeting it feeds.GTM AI Academy on YouTube What you can do after it
- Connect the tools you already pay for to Claude, so the work moves into one window instead of 15 tabs.
- Write a scheduled task with a wide prompt: what the product does, where it is going, and the data it may read.
- Keep few jobs. Each one needs an owner and a meeting or action it feeds, or it goes.
- Sort your reuse. A one-off is a prompt. A repeated job with context is a project. A way of doing something is a skill. A bundle of those with a connector is a plugin.
- Send the result to the whole team. A job everyone reads turns AI from single player into multiplayer, in his words.
- Review every output. He says it is not perfect and still has to sound like you.
The path through it
- 7:06The connector stack: email, calendar, call recorders, CRM, Notion, Slack and a few hand-built MCP servers.
- 8:07Scheduled tasks: the 4 p.m. outreach pulse and the end-of-day follow-up audit.
- 9:00The product-market-fit review that runs after every 20 prospect calls.
- 10:00What that job needs: a wide prompt, the call recordings and the company data.
- 11:30His claim that a direct CRM connector reads about 30 percent of what it needs.
- 18:15The three mistakes, including 200 jobs nobody read.
- 20:17Two change-management rules: fewer jobs, and every job has an owner and a workflow.
- 22:20Artifacts as shared dashboards.
- 23:20Prompt, project, skill or plugin: how he decides.
- 25:20A free library of sales prompts he published.
- 29:26His advice: play with it, give context, review the output, ask for short answers.
The risks
Named in the video
- Too many agents. With 200 scheduled jobs nobody read any of them; 'less is more'.
- Information without action. A job with no owner and no meeting attached is noise and gets cut.
- Output that needs review. He says to check everything it writes.
- The context window. A model connected straight to a messy CRM samples rather than reads, by his estimate about 30 percent of what matters.
Left out, so AI Lately adds them
- He sells the fix. BackEngine is the data layer he says solves the 30 percent problem. The 30 percent is his estimate, not a measurement shown in the video.
- The people on the calls. Connectors read a whole team's inboxes, calendars and call transcripts. The video does not say what the people on those calls agreed to.
- Cost. Scheduled tasks and connectors run on paid Claude plans. No spend is named.
AI Lately's test
- Can a beginner follow it?
- Partly. The jobs and the rules are easy to copy. The screen share moves fast, and the connector setup is assumed rather than shown.
- Does it cover the risks?
- Partly. He is candid about his own mistakes and the review rule. The vendor interest and the consent of people on recorded calls go unmentioned.
Portnoy's company, BackEngine, sells the data layer he recommends in the video. The host says he worked on a similar product at Momentum.
Build a Clay table that finds the sales leaders who care about developing their teams
Clay Tutorial: How to Find B2B Buyers Who Actually Care
Arup Chakravarti shares a live Clay table of UK healthcare providers. Prebuilt AI columns sort each company as B2B or B2C and pull its last ten articles into themes. Another column scores its strategic priorities green, amber or red. A second table lists the sales leaders at those companies. A prompt column reads each leader's LinkedIn posts and comments and scores them a strong, moderate or weak advocate for professional development. Clay then drafts an outreach email from the whole stack.GTM AI Academy on YouTube What you can do after it
- Tell a B2B company from a B2C one with Clay's prebuilt AI columns. They carry the Clay logo in a hat.
- Pull a company's last ten articles as JSON, then split the titles, bodies and sources into their own columns.
- Write a structured prompt column: the context, the objective, the fields it reads, and a three-step scale with a reason for each score.
- Widen a prompt that comes back too narrow. 'Advocate for the sales function' found few people; 'advocate for professional development' found many more.
- Spot-check company sizes against annual reports. Clay came closer than LinkedIn, which counts resellers and influencers as staff.
The path through it
- 6:50What Clay is: an online spreadsheet that runs AI and bought data on every row, paid for in monthly credits.
- 9:35The Google Maps integration finds local firms that are not on LinkedIn at all.
- 11:40The use case: UK healthcare providers, from a sales-training case study.
- 18:10The live table: the prebuilt B2B or B2C column, growth mode, and the ten-article theme pull.
- 22:30Strategic priorities scored green, amber or red by confidence.
- 23:35The sales leaders sub-table, and how he spot-checks the data.
- 25:30The advocacy column: a structured prompt that scores each leader strong, moderate or weak, with a reason.
- 30:25Clay drafts the outreach email from every signal in the table.
- 31:40Results so far: none. He has not sent any of the emails.
- 32:30How hard Clay is to learn: 'a little fiddly', and the free Clay University.
The risks
Named in the video
- Profiling people from their posts, comments and likes. He says LinkedIn is turning political and the same method could soon sort people by politics, which he calls 'possibly dangerous'.
- Bad data. Spot-check company sizes against published reports; LinkedIn's counts run high.
- Volume. The same table can send outreach straight from Clay, so it makes spam faster as easily as it makes outreach sharper.
- No results yet. He says plainly that he has not operationalized any of it.
Left out, so AI Lately adds them
- Consent and the law. The table profiles named people from their LinkedIn activity. The video never mentions GDPR, which binds a UK list like this one, or LinkedIn's terms on scraping.
- Cost. Clay runs on monthly credits and every AI column and data pull spends them. The video names no prices.
- Checking the scores. A 'strong advocate' label is a model's reading of someone's posts. The video shows one reason text and no check of how often the label is right.
AI Lately's test
- Can a beginner follow it?
- Partly. He says up front that he assumes you have already built a table. He also had to learn JSON to pull fields out of the article data. Take Clay's free course first, then watch this.
- Does it cover the risks?
- Partly. He raises profiling and bad data himself. He skips consent, the law and cost.