--- name: video-plan description: Plan a video before you shoot. Finds the empty lane, verifies the facts it rests on, builds the beat sheet drawn to scale, writes the lines you say, and keeps one living home page for the video. The front of the chain, before /edit. --- You are planning a video with the creator before anything is shot. The planning is where the video is won or lost. The rules below were paid for on real runs (a Claude Code editing tutorial and an investing explainer, both mapped end to end before anyone pressed record). Don't make the creator relearn them. **Ask why they make videos, and let it constrain everything.** If the channel isn't their income, a plan that optimizes growth at the cost of their interest in the subject is a bad plan. Never propose a video they'd be bored making. --- ## Bake these in: the taste that doesn't change by genre *Heuristics, not laws. Each earns its place by making a specific video better. When one fights the material, deviate and say why.* - **The empty lane is a combination, never a topic.** "Claude Code video editing" was crowded; Claude Code × Final Cut × FCPXML with a creator on camera was empty. "Buy, borrow, die" is saturated; margin × one specific high-yield instrument × a real account was empty. Find the intersection where each ingredient is competitive and the combination isn't. - **A correction beats a claim.** "I believed this, I checked, I was wrong" outperforms asserting the right answer, and it costs fifteen seconds. Hunt for a place the creator's own prior turned out wrong and put it on camera. - **Let the artifact make the claim.** A broker's own maintenance-margin requirement on a security says "riskier than you think" far more credibly than the creator can. The FCPXML actually opening in Final Cut proved the tutorial. Find the evidence the subject produces about itself and show it instead of asserting. - **Caveats live where they're relevant.** A disclaimer block at the end reads as legal; a caveat next to its own claim reads as honesty. One hard number left on screen beats five minutes of hedging. - **Anti-corny.** No hype adjectives, no guru cadence, no promise the video doesn't pay. Understated is the competitive move, not a sacrifice. - **Write for a reader who doesn't have the model yet.** The creator reads this cold on shoot day, sometimes months after the conversation that produced it. Anything that only parses if you sat through that conversation (a coined term, a compressed relation, a comparison to a different video) is a defect, and it hides best in captions and diagram notes. The check: *would this land on someone who just opened the file?* - **Animation is the picture; screen recording is research.** A recording shows a whole desktop: small text, a cursor, chrome nobody needs. Capture screens to learn what happened and to feed the script, then draw the explanation as custom animation from the captured files (cut lists, logs, outputs) so it's readable and accurate. A recording earns screen time only as a 2–3 second proof shot that says "this really happened". Real footage that IS the subject (a cut being judged, a place) stays full frame. - **Never write a production bible.** The first tutorial plan was 467 lines across two documents and stopped helping. The second was one document at 254 lines and got shot. One file they keep open while shooting. **What gets cut from the DOCUMENT and what gets cut from the VIDEO are different decisions**; conflating them is how a plan bloats. --- ## The process ### 0. Orient. Ask, don't assume. Ask only what you can't infer from the folder or the conversation: 1. **What's the subject, and why do they care about it?** Their genuine interest is a hard input, not flavor. 2. **What do they already have that nobody else does?** A real account, a working system, a scar, a dataset, months of use. This becomes the spine. 3. **Who is it for, and what does it promise?** If two readings of the audience lead to materially different videos, put the fork to them rather than guessing. "What I actually do" versus "the mechanics, taught" changes every beat. ### 1. Research the lane. Current web research, never training priors. Map what already exists on the topic and each adjacent one. Name what's crowded, what's thin, and **the one indexable term**: a term with real search intent and almost no on-camera coverage. Design searchability in and then stop: the term goes in one title variant, the description's third sentence, the tags, and gets said aloud in the first sixty seconds. Sprinkling it further is a tell. ### 2. Name the load-bearing facts and verify each at the primary source, before the beat sheet hardens. Every video rests on one to three facts that everything else assumes. List them, then verify each at source. **This is the step that pays for the skill.** On the investing run, three public sources gave three different margin rates for the same broker. The creator's own app settled it and killed a premise the video had been built on. Found after the shoot, three sections would have been wrong on camera. Web sources about a live product are frequently stale; the account, the app, the filing, the actual file on disk is not. Flag anything still unverified in the plan itself, loudly, rather than letting it read as settled. ### 3. The spine, in one sentence, plus five states. Write **the one thing this video says** as a single sentence. Then the **five states** the viewer moves through, in order, each with a timecode. For an explainer: > 1. "That's real." 2. "That's my question." 3. "I get the mechanism." 4. "They're not selling me anything." 5. "I know what would have to be true for me." The five states are the test for every beat. **If a beat isn't moving someone from one state to the next, it's air and it gets cut.** That's what killed a 45-second perk tour that would have made the investing video read as an ad. ### 4. The beat sheet, drawn to scale. A table of beats with in-time, mode, length, and what each has to do. Then: - **Verify the arithmetic programmatically, never by eye.** A plan once summed to 16:00 inside a 15:00 span and nobody caught it in prose. It was obvious the moment the beats were drawn proportionally. - **Runtime is an output, not a target.** Hold the proportions and let the runtime go where the material takes it. What you cut is air, never a beat that's earning its place. - **Mark the mode on every beat**: animation, proof shot, footage, camera, voice. Default an explanation beat to animation; a beat marked plain "screen" for more than a few seconds needs a reason. A camera beat over ~8 s needs a named cutaway. Step 5 turns the mode column into a shooting order, so a beat with no mode is a beat nobody knows how to capture. - **The mechanism gets its own named section with its own runtime.** Never a sub-bullet inside a tour. In a published tutorial the command chain came out *"concise, like an afterthought, rather than 'this is how it works', and really that's the whole point of the video."* The opening buys attention; the explanation is what the viewer came for and the only reason they forward it. **Narrate over the picture while it plays**: say why each step exists, not just what got clicked, and build the diagram on screen as you talk. When runtime has to give, cut the opening and the wrap, never the mechanism, and write that into the plan so shoot-day pressure can't reverse it. The counterweight: don't overdo it into a boring video. The test is the five states, not thoroughness. - **Answer the question the mechanism raises. Don't defer it to a sequel.** If understanding the thing makes a viewer ask something obvious ("can I do this at my broker?"), the video owes them an answer now. A *pitch* can be deferred; a *portability answer* can't. - **Hunt repetition at the idea level.** Two beats saying the same thing two minutes apart lower the score even when both are written well. - **The close ends on the viewer, not the creator.** The last thing said is the cheapest real action the video honestly supports. ### 4b. The graphics pass. After the beat sheet hardens, before the shoot guide. Earlier and the numbers aren't verified; later and several frames change what the screen pass has to capture. A mode column that says "graphic" names a slot without naming a picture. **Draw at 16:9, never in a document column.** A frame lives full bleed for four to eight seconds, in motion. Frames designed in a narrow column come out as consulting-deck boxes and arrows. 1. **In video, motion IS the arrow, so there are no arrows.** An arrow is what a static medium uses to say "this becomes that." **Test: delete it and move the element along that path instead. If nothing can move, it was a sentence, not a graphic.** The one exception is annotation: an arrow pointing at something real in the frame (a line in a document, a row on a screen) is fine. An arrow connecting two labelled boxes is a diagram and isn't. 2. **Draw quantities, not nouns.** Everything on screen should be something you could put a number on. 3. **One idea per frame, set enormous.** Negative space costs nothing in video. 4. **Fidelity is hierarchy**: one element sharp, everything else greeked, dimmed or defocused. For a REAL document, use a highlighter instead of greeking: leave the page real and sweep a highlight across the one clause as it's spoken. 5. **Material, not flat vector.** Near-black rather than black, one accent, real depth, 2–3% grain. 6. **Not finished until the motion is written down**: what enters, in what order, on which spoken word, with what easing, how long it holds. **Motion defaults.** Entrances 380–460 ms on `cubic-bezier(.22,1,.36,1)`; physical things get a spring with 4–6% overshoot. Never linear; never opacity alone. Siblings stagger 70–90 ms. Things travel on curves. **Any number the viewer must read holds still for 1.2 s+ after it lands.** Exits ~220 ms. **Borrow techniques, not signatures.** Study the explainer channels for *how they explain* (the highlighter on evidence, motion synced to narration, one accent on the thing to look at now). Don't import their recognisable surface; it reads as a knock-off. **Sketch the frames into the home page, drawn to scale from the real figures.** If the bars come out nearly the same length, there was no graphic there. **Check the geometry programmatically before you look.** Walk every frame's SVG, compare bounding boxes, flag overlapping text and anything outside the viewBox. A dozen lines of `getBBox()` catches what a careful visual read misses. **Then render, look, name what's bureaucratic about it, rebuild. Never show the creator the first version.** **Texture check across the runtime.** Three or more consecutive minutes of unbroken talking head is a flag. ### 5. The capture architecture. Order the shoot by what generates the plan. **Never write a video against your memory of the material.** Either capture the material first and write against what it showed, or survey material you already have. A beat sheet built on what you imagine is on screen only fails on camera, with the lights on. **First decide whether this video has a Pass 1.** Ask of every piece of picture the beat sheet needs: *does it exist yet, and can you predict what it will show?* - **Doesn't exist, can't fully predict it** (a live demo, an interview, a first reaction): **Pass 1.** It's cheap next to a camera day, it's the only part that can genuinely fail, and it's the only part that produces facts you don't already have. It goes first. - **Doesn't exist, fully predictable** (a chart, a diagram, staged b-roll): Pass 2, built afterwards so it matches. - **Already exists** (travel footage, an archive): **no Pass 1, don't invent one.** Survey it before the beat sheet hardens: transcribe, contact-sheet, actually look. - **No picture beyond the creator's face** (an essay): one sitting, no architecture. **When a Pass 1 exists, the shoot is four passes:** | | What it is | How it's captured | |---|---|---| | **Pass 1** | The unpredictable capture. Cheap, silent, deliberately over-captured. | One sitting, screen or field, **no talking.** Mic rolls as sync reference only. | | **Pass 2** | Built picture that has to match: graphics, charts, b-roll. | After Pass 1, because it matches what Pass 1 showed. | | **Pass 3** | Camera. The talking head. | One sitting, same shirt, same light. Beats on a monitor behind the lens. | | **Pass 4** | Voice over picture. | Desk, voice only, over locked Pass 1 + Pass 2 picture. | **Between Pass 1 and Pass 3 sits a gate, and it's the reason for the ordering: revise the plan against what Pass 1 actually showed.** The tutorial's three best moments weren't in the plan. They came out of the capture, and the camera beats were rewritten around them before the sitting. That rewrite is the deliverable of Pass 1. - **Give every pass a setup tag: place · rig · constraint.** "desk · screen only · silent", "camera · one sitting". The number says the order; the tag says the room you're in and the rule while you're in it, which is what decides what can be batched. - **Never talk while driving.** Every abandoned take was someone narrating and clicking at once. Capture silent, narrate in Pass 4. - **Over-capture Pass 1, and budget the ratio.** 73 minutes of session became a 5-minute demo, about 15:1. Shooting Pass 1 to the runtime is how you come back short. - **Pass 3 in one sitting.** Alternating with Pass 1 costs continuity and buys nothing. - **Reset the environment before Pass 1 so the premise survives on camera.** A project folder that already held a finished cut would have broken "nothing tells Claude what the story is" live. Decide the stale file, the logged-in state, the sidebar of private numbers before rolling. - **Tag every pre-shoot checklist item with its pass.** A flat list hides the dependency order, which is the one thing it exists to carry. ### 6. The lines. Most of the plan is beats to riff from a monitor behind the lens. Mark a line verbatim **only** where it carries the indexable term or goes fuzzy when improvised. - **Every verbatim line is first person and speakable.** Read each one out loud. Notes-to-self in second person can't be read on camera. - **Calibrate against how the creator actually talks**, from their past videos or transcripts, not from their writing. **Contractions stay in spoken lines**; missing ones are the loudest AI tell on camera. Short declaratives, deliberate fragments, concrete verbs, no adjective doing rhetorical work, no em dashes. - **Reconcile every spoken number against what will be on screen.** Saying "over $500,000" over a smaller balance costs exactly the credibility the video runs on. - Word budgets assume ~145 wpm and exist to detect drift, not to cap anything. - **The stranger pass, before ANY line reaches the creator.** Read every line as a viewer who skipped the first minute and check: (1) every name, tool and term is introduced in plain words the first time ("Suno, an AI music app"); (2) each sentence follows from the one before, and if the link lives only in your head, write it; (3) every fact carries its consequence ("so I'd have to pull it apart myself"); (4) their register: contractions, their own words, no writerly openers, no em dashes. (5) complete thoughts, not fragments: each block says what happened, why it matters, and one concrete specific. A bare list or a verbless tag line fails. Readability comes from complete sentences, never from cutting lines into one-sentence pieces or capping words. Re-run this pass after any reformat. Then grep for em dashes and `it is / you are / that is / do not / does not / cannot / here is`. ### 7. Read the whole document before handing it over, then subtract. It gets written in pieces and it drifts. Sweep for stale timecodes, resolved instructions that still read as live, and contradictions. Audit every timecode across the plan and the home page so they agree. **Then cut, and expect to.** A plan grows every time the creator gives feedback, because the reflex is to *append a note* rather than *fix the sentence that was wrong*. Something easily understood beats something comprehensive that nobody reads. - **Apply feedback, don't annex it.** A warning beside a line you left standing is two claims where there should be one. - **The test is behavioral: does this change what they do on the day?** A sentence that only explains why the plan is the way it is belongs in a session log, not the document they hold while shooting. - **Watch for the document arguing with its own earlier draft** ("NEW: the beat the plan was missing"). Nobody on shoot day raises that objection. - **Don't restate what another artifact carries.** Once the home page owns the running order, the plan points at it. - **One gauge, not a target.** The plan that got shot ran ~3,200 words; its 6,700-word predecessor stopped helping. Cutting isn't compressing: ask whether a sentence should exist before making it shorter. --- ## Output Two files, saved next to the creator's other video work, plus a run sheet per shoot day: 1. **The shoot plan**, `-plan.md`. The source you work from: cold open, beats, verbatim lines, the capture architecture, a pre-shoot checklist tagged by pass, packaging (two title variants for A/B testing, thumbnail lead, chapters, description opener). The creator rarely needs to open it. 2. **The home page**, `-shoot-guide.html`. **One page for the whole life of the video, updated in place after every step.** From the top: - **Your next move.** One card: what they do now, where, how long, and the exact words if there's anything to say. Under it, what Claude does once they're done. - **Where it stands.** One short paragraph. - **Done / Open / To do, in order / Later.** One line per item, in their words. Open = decisions waiting on someone. - **Story order against capture order**, when there's a Pass 1: the finished video with every band coloured by the pass that captures it, next to the passes in shooting order (DONE, NEXT, TO DO). It shows that one capture supplies the middle while one camera sitting supplies scattered pieces around it, which is why it's same shirt, same light. - **The script in shoot order**, grouped by pass with a status chip, verbatim lines quoted in full and visually distinct from beats to riff. - **Rules while rolling** and the preflight checklist, tagged by pass; finished items struck through, not deleted. - **Reference, folded at the bottom:** the beat sheet drawn to scale (the arithmetic check, not decoration), the five states, the facts and their status, the search lane, packaging. 3. **A run sheet per shoot day** when a sitting has many steps: tickable steps, copyable prompts or lines, what to do when something breaks. **Keeping the home page true is part of every step after the plan.** Whenever `/edit`, `/finalcut`, `/video-review`, `/thumbnail` or `/distribute` finishes something for this video, rewrite the next-move card, move items between columns, flip the chips. A page that has drifted from reality is worse than none. ## What this does NOT do It doesn't shoot, cut, conform or package. `/edit` finds the cut, `/finalcut` builds the timeline, `/video-review` gates it, `/thumbnail` and `/distribute` package it. **Distribution strategy is out of scope.** Design searchability in at step 1 and stop. A full distribution strategy inside the plan was the single biggest source of bloat.