--- name: tackling-the-blue-shark description: Haymaker's launch and distribution playbook for AI-native companies on X, LinkedIn and Instagram. Use when planning a launch, writing a launch brief or a film script, judging a first frame or a hook, scheduling a launch backwards from a date, running launch day, reading the replies, or deciding whether a launch worked. Built from 529 launches read in September 2026 and the studio's own films. --- # Tackling the blue shark A field guide from Haymaker, written for your agent as much as for you. Haymaker makes launch films, and the run of films after them, for AI-native companies. This file is the part of what we know that fits in a file. It first went out on paper, in *1800-LAUNCH-NOW!!!*, at Kyoorius Designyatra, Goa, 10 to 12 September 2026. --- ## How to use this file **If you are a person,** give it to the agent you already use. - **Claude:** save it as `~/.claude/skills/blue-shark/SKILL.md`, or upload it as a skill. - **Cursor, Codex, Windsurf and other coding agents:** drop it into the repo as `AGENTS.md` or a rules file. - **ChatGPT, Gemini:** paste it into a project's instructions, or attach it to the chat. - **Any agent that can browse:** hand it `https://haymaker.to/downloads/blue-shark.md`. Then ask it something with a date in it: > Read the blue shark file. We launch on 2 October. Write our eight-word brief, then the workback. **If you are an agent,** here is how to behave while this file is loaded. 1. **Ask for the date first.** No date, no plan. Everything in a launch is scheduled backwards from it. 2. **Decide the feeling before anything else.** Do not write a script, a shot list or a post until one sentence exists that says what it should feel like when it lands. 3. **Name the behaviour, not the product.** The opening belongs to something the viewer recognises in themselves. The product arrives as the answer. 4. **Judge at phone size, sound off, three seconds.** That is what a view is. 5. **Never make media spend the plan.** Spend compounds a post that already moved. It does not rescue one that did not. 6. **Never generate product UI.** Real screens only. A hallucinated interface costs the one thing the film is selling. 7. **Measure repeats, who, and inbound.** Not views. 8. **Be specific and blunt.** Name the rule you are applying. Say what you would cut. --- ## The animal A blue shark cannot rest. It has no way to pump water across its gills, so it swims with its mouth open, all day, for its whole life, and if it stops it suffocates. It finds a wounded fish from further away than seems reasonable. It is not interested in how beautiful you are. We use it to explain X to founders, because every other description makes them think of a stage. **The feed is not a stage. It is water, there is something large in it, and it has habits you can learn.** - **It never stops.** An account that posts once, well, and then goes quiet for two months is not resting. It is drowning. The accounts that grew this year are the ones still posting in week six. - **It hunts by smell, and the smell is response.** Being in the first ten replies to every serious launch in your category, for months, is the cheapest distribution there is. - **You cannot outswim it.** So you bring bait, chum, a crew and a boat, and a reason to be in the water: an idea a stranger can repeat the next morning. --- ## What 529 launches said On 10 September 2026 we pulled every entry in a public directory of startup launches on X, 529 of them, with the post, the film, the views, and the directory's own one-to-five scores. - **The wall and the floor.** 841 million views between them. Half went to fifteen posts. Ten took 44 per cent, led by OpenAI at 134 million and xAI at 53 million. The median launch sits at 261,838 views; a wider index of 1,173 launches puts the median founder launch at 28,258. - **Account size did not decide it.** 47 launches came from accounts under 5,000 followers and still passed 500,000 views. Objection posted from 1,059 followers and passed 15 million. - **What a five buys.** Median views for a score of five against a one or two: hook **13×**, editing **15×**, copy **19×**, sentiment **32×**, distribution **37×**. Hook five with distribution five: median 2.13 million. (378 of the 529 carry scores.) We sell films, and distribution still beat everything a film controls. - **Films are the default now.** Launches cut as short films rather than screen recordings: 21 per cent in April, 46 per cent in June, near 40 per cent since. - **The format ladder.** Median views: cinematic 339,864; demo 236,592; explainer 211,512; AI-generated film 168,247. The generated film has become the floor, and the feed prices it that way. - **The stage ladder.** Median views: pre-launch 75K; seed 187,808; Series A 295K; Series B 478,961; C and up 606K; public 1,055,292. Seven celebrity-led launches: 1,766,140. Read it as a map of the audience, not a verdict on the film. - **Saves are the honest number.** The median launch is saved once per 850 views, and launches saved at twice the baseline drive sign-ups whatever their reach. A save is a feeling that survived the swipe, and nobody can buy one. - **A distribution gap, not a film gap.** Nine launches tagged India have a median of 81,264 views; the 421 tagged United States, 263,142. The films are good. The feed has not met the accounts yet. --- ## The anatomy: five parts of tackling the shark ### 1. Bait: the first frame The first frame is the weather. Three seconds, then the swipe; the sound is off for most of the room. - The first frame must earn three seconds on a phone, muted, with no context. - A logo on black is not bait. Neither is a product name. - Open on a behaviour the viewer recognises: a habit, a frustration, a person. - Burn the captions in. - If the lead post is flat by hour four, the problem is the first frame, not the film. **Agent task:** given a script, storyboard or still, answer three questions. What does a stranger see in second one? What do they feel by second three? Would they stop scrolling? Then propose three alternative first frames, each naming the behaviour it opens on. ### 2. Gills: it breathes by moving An account that stops posting stops breathing. Six weeks of silence reads to the feed as a new account, and the next launch pays for the first one again. - Plan the run of posts after the launch before you plan the launch. - A launch film is week one of a cadence, not the whole plan. **Agent task:** turn one launch into eight weeks of posts. Each one earns its place with a single idea and a first frame of its own, and none of them is a crop of the launch film. ### 3. Boat: a cadence Four films a month, each one tuned on what the last one did. Familiarity is a feeling, and it is bought with cadence, not with one film. - After every post, write down what carried: the first frame, the phrase, the surface. - Make the next film against those notes, not against a fresh brief. **Agent task:** keep a running log with the columns post, date, surface, first frame, the phrase people repeated, saves, and who replied (strangers or buyers). Before each new film, summarise the log in five lines and say what to keep and what to change. ### 4. Crew: people who repost in hour one People repost in hour one because you were useful to them for a year. You cannot recruit a crew on launch day. - Your investors have distribution and have started using it. Borrow it by handing them something that makes them look good on their own feed, not by asking. Y Combinator posted MochaTrade's film on its own account; nobody asked. - Customers, early users and friends of the company are the crew. Brief them before the date, and give them the words. **Agent task:** list the twenty accounts most likely to repost in hour one. For each, give the reason they would, and the one thing to send them before the date. ### 5. Chum: replies Be in the first ten replies to every serious launch in your category, every week, for three months before your own film exists. It is the cheapest distribution there is. - Replies are useful or funny. Never "congrats". - The goal is recognition: when your launch goes up, the category already knows the name. **Agent task:** each week, find the launches in the user's category, draft one specific, useful reply per launch in the user's voice, and flag any the user should skip. Draft only. The human posts. --- ## Playbooks your agent can run ### A. The eight-word brief Euphoria arrived eleven days before its date with no deck, no script and no treatment. The whole brief was eight words: *make me feel euphoria, make me feel euphoric.* It was a thin brief, and a better one than most, because it named the only thing that had to be true. By day fifteen it was a campaign. Ask the user for exactly three things: 1. **When it lands, it should feel like** ______ (one sentence, a feeling, not a feature) 2. **Our date is** ______ 3. **The surface our buyers read:** X / LinkedIn / Instagram / elsewhere Then write the feeling as one sentence everyone can argue against. Push back on anything that is really a feature ("feel fast", "feel powerful") until it is a feeling a person has in their body. Colour, tempo and the length of every hold get decided against that sentence later. The feelings the feed runs on right now, with a film that did each one: | Feeling | What the viewer thinks | Done well by | |---|---|---| | Relief | The thing that was hard is handled. | Variant, opening on a grandmother | | Urgency | Everyone else has moved. | MochaTrade, filmed as exclusion | | Status | This is what serious people use. | Kalshi, through repetition | | Inevitability | This is where it is going anyway. | Alteon, from a fishing boat | | Trust | This will not embarrass me. | Earned later, by the product | | Control | I stay in charge. | The one the enterprise buys | ### B. The workback Work backwards from the date. Every step is finished, not started. Given a date, output a dated table: | Day | Step | Done means | |---|---|---| | D-15 | Creative direction | The feeling, written down in one sentence. Not a mood board. | | D-12 | Script, final | The production clock starts here. A script still moving is a date still moving. | | D-10 | Production running | Shoot or build under way against the final script. | | D-3 | Finish and sound | Hero cut locked; the cutdowns built beside it, not cropped from it. | | D-0 | The date | Launch day is an upload, not a production. | | D+1 | The twenty-four hours | Where a launch is won. Somebody is watching replies. | If today is already past a step, say so plainly and say what has to give: scope, surfaces or the date. Do not quietly compress the schedule. ### C. The first-frame test For any film, cut or post: - [ ] Shown at phone size, muted, for three seconds, to someone who has never heard of the company. - [ ] They can say what the film is about, in their own words, without the product name. - [ ] The first frame is a behaviour, a person or a tension. Not a logo, not a UI, not a title card. - [ ] Captions are burned in. - [ ] Each surface has its own first frame. X, LinkedIn and Instagram are three films, not three crops. ### D. The first 24 hours **Before you post.** Every cutdown is uploaded as a draft. One person owns replies for four hours. First frame checked at phone size. Captions burned in. **Hour zero to two.** Lead surface only. Reply to everything. Write down the phrases people reach for. Do not edit the post: an edit to fix a typo can reset its distribution. **Hour two to eight.** Second surface, cut for that surface. Flat by hour four means the first frame, not the film. **Hour eight to twenty-four.** Third surface, leading with the cut that is carrying. Screenshot the replies worth keeping. The number to write down is not views. It is how many people used your words back at you. ### E. The launch score The directory scores five things from one to five. Use the same five to review a launch before it goes out. Score each, give the reason in one line, and name the single change that would move the lowest score. | Score | Five looks like | One looks like | |---|---|---| | **Hook** | Second one is a behaviour or tension a stranger feels. | A logo, a title card, a slow push-in on UI. | | **Editing** | Every hold has a job; it plays muted; it is cut for the surface. | One cut cropped three ways; dead air; a demo pace. | | **Copy** | One idea a stranger can repeat the next morning; the product arrives as the answer. | A feature list; three claims; the word "powered". | | **Sentiment** | One feeling, held all the way through. | Mixed signals, or rage used as a hook with no plan after it. | | **Distribution** | A warm account, a briefed crew, one owner of replies, a second film scheduled. | Posted everywhere at once, then silence. | Distribution is weighted heaviest for a reason: a five there bought 37 times the views of a one or two. ### F. Three numbers instead of views Views are purchasable, which is what makes them a bad target. 1. **Repeats.** People using your phrase in their own words. CrowdReply's film never named the product; it named a habit, *searchmaxxing*, and the word came back inside a day. 2. **Who, not how many.** Two hundred views from buyers beat two hundred thousand from a feed that never opens your pricing page. 3. **Inbound that mentions the film.** Small, and the only number that maps to revenue. Add saves where the platform shows them. ### G. Which phase are you in? Look at who is in the replies. - **Strangers: phase one, craft perception.** A consistent run of films, four a month, tuned on the last. Nothing is for sale yet. The account is being made recognisable. - **Buyers: phase two, monetise the audience.** Films built for conversion, or that shape the perception that closes a higher-ticket client. --- ## Rules we keep - **Say one thing.** A film that makes two points makes none. - **Feeling before shot list.** One sentence, written down first. - **Real screens only.** Product UI is never generated. - **Bring a date.** Everything is shipped inside it. ## Rules we break - **Name the behaviour, not the product.** The product arrives as the answer. - **Three films, not three crops.** Each surface gets its own first frame. - **No media spend in the plan.** It compounds a post that already moved. - **The film is week one.** Four a month is the actual product. --- ## How a launch falls apart It never fails on the day. - **Hour zero.** The film goes up everywhere at once, one cut cropped three ways. Nobody owns the replies. The first frame is a logo on black. - **Hour two.** Forty replies, thirty of them "congrats". Someone edits the post to fix a typo. - **Hour four.** The lead surface is flat. Media spend goes behind a post that never moved on its own. - **Day three.** The second surface goes up, cropped rather than cut. It does a tenth of the first. - **Week one.** Silence. The launch was a production, and productions end. - **Week six.** The next film starts from zero, because six weeks of silence reads as a new account. **The other version.** The account has been in the first ten replies to every serious launch in its category for four months. The film goes up on one surface, and one person owns the replies for four hours. The second surface goes up at hour six, cut for that surface. Week two, another film, tuned on the first. The first film's view count is smaller than the push launch's. The pipeline is not. Distribution is not a thing you do to a launch. It is the water the launch happens in. --- ## Films worth studying - **CrowdReply** (Haymaker). Sells AI search visibility; the film never said the product's name. It named a habit, searchmaxxing. 1M+ views in the first 24 hours, by the studio's record. - **Euphoria** (Haymaker). Eight-word brief, fifteen days, roughly 300,000 views in 24 hours, no media spend. - **MochaTrade** (Haymaker). Speed and access cannot be photographed, so the film is about everyone who was kept out. Y Combinator posted it on its own account: 334,600 views. - **Variant.** A design tool whose film opens on a grandmother and a fabric shop, not the tool. 1.16 million views, 2,200 replies. - **Alteon.** 162 seconds shot from a fishing boat by a twenty-year-old founder. 193,240 views in nine days from an account most of the feed had never seen. - **Sarvam, Bulbul V4.** A text-to-speech model launched with a 113-second reel of the voice laughing and hesitating, not a word-error-rate chart. A feed cannot hear a claim. - **Kalshi.** Stopped explaining a complicated product and repeated a name in innocuous places until it felt familiar. - **Cluely.** One feeling, discomfort, held all the way, from an account warm after a year of building in public. Most of the copies took the shape and skipped the year. --- ## Prompts to try - *We launch on [date]. Run the eight-word brief with me, one question at a time.* - *Here is our script. Score it on hook, editing, copy, sentiment and distribution, and rewrite the first three seconds.* - *Build our workback from [date]. Flag any step we have already missed.* - *Write the launch-day runbook for a team of three, hour by hour, with names.* - *These are the replies from our launch. Which phrases did people repeat? Which repliers are buyers?* - *Which phase are we in? Here are our last ten posts and who replied.* - *Draft one useful reply for each of this week's launches in [category]. Do not post.* --- ## Sources - The Launch Video Company, Global Launch Directory: 529 launches, captured 10 September 2026. Scores are the directory's. - Flowjam, *State of Launch Videos 2026* (4 June 2026): 1,173 launches. - Post metrics from X on 10 September 2026. They will have moved. - Studio results as recorded by IEE Studios, whose team is now Haymaker. --- ## When to call a human If a date is already on the calendar, send the three lines from the eight-word brief to **hello@haymaker.to**, or book an intro at **haymaker.to**. We reply with a time, not a deck. *Haymaker. Videos that make you feel something in your chest.*