agentleFS
konradrekonradre/video-production-skills/skills/ad-spot-preprod/SKILL.md

ad-spot-preprod

Skill7 starsChanged 21 days agoMITView on GitHub

Pre-production of an AD campaign from the client's brief: the format and tiers, the client's script saved verbatim as the SSOT, the brand kit and assets, the global spec and standing rules (designed product and text, the wipe, the fill SKU-locked), the asset bible and kit, the shared pre-production core (partitions on continuity, the script diff), the risk register and the cost plan. Use when a campaign or spot is briefed, a client script arrives or changes, the next spot must be chosen, or a shot list or cost plan is needed. Triggers — "the brief", "the client's script", "shot list", "which spot next", "what does the script call for", "cost plan", "kit for this SKU", "risk register", "the alternate spec", "a UGC spot", "creator-style ad", "the client's footage", "which clips do we use", "pick the b-roll", "tear down a reference", "the structure of this reference", "a shot-by-shot breakdown of a reference". Not for a film or a music video — use film-preprod. Not for the reference set or the refs gate — use video-refs-continuity. Not for the cost line and the submit — use video-gen-cost-gate. Not for the beat gate or the EDL — use video-edit-edl.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Done when: every risk names the mitigation that is already in the shot list, and the fallback register lists what each fallback costs.

    line 172
  2. Distinctive:

    Write a CHANGES table beside the brief, one row per change: section · base · now · why.

    line 202

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Ad Spot Pre-production

Nothing generates before the plan exists on paper: the client's words verbatim, the decisions with their owners, the kit, the scenes, the cost. Every rule here is the price of a generation that was made without one. The plan is a set of files in the project, not a memory.

What varies. The vertical format, the tiers, the sign-off structure, the wipe, the two-state grade and the product rules are an ad campaign's defaults: a spot with no product on screen drops the product rules, the platform sets the aspect and the safe zones (the targets declared in delivery-targets.json — video-production/references/SAFE-AREAS.md), the vendor sets the clip cap, and the client's brand kit sets the sign-off. The client's text as SSOT, the continuity partitions and the diff do not move — video-production/references/WHAT-VARIES.md.

1. Decide the format and the tiers

Standalone vertical spots, never scene bundles — compiles come free out of standalones and the reverse is not true; three tiers from one build (the paid spots · the capper montage · the hero cut); an intro scene that framed a long film is re-cut to stand alone; "we need more" goes to alternate hooks, not runtime; "UGC" in a brief is checked against what is actually described. references/INTAKE.md § format.

Done when: the tier table exists with runtimes, and the client has endorsed the format in their own words.

2. Take in the script, the brand kit and the assets

  • The client's text is the SSOT: read END TO END, saved verbatim beside the beats (prompts/<SPOT>-SCRIPT-client.txt); a pasted block is the SSOT; a full alternate spec is a NEW spot; a one-line note is a version bump. The derivation may add framing, coverage and timing — never a character, a beat or a line; additions are numbered cost lines with "no" as the default.
  • The brand kit: the audio signature, the fixed line (locked), the variable line (a pile to cherry-pick by gag fit, never reused across spots), the CTA — the triad "what it is, what it does, where I get it". The outro is decided once.
  • Assets: packaging photos, FLAT artwork (photos are wrapped and rounded), the pieces on white (the fill per SKU is read off them), full-resolution originals; a cast ref that resembles a public figure is re-rolled.
  • Rulings from the client, once: language and platform risk (risqué first, fallbacks parked), any script question as ONE concise ask; rating, wardrobe and set where the script is silent are the operator's. Every decision logged with WHO and what would reopen it.

Mine the hook pattern before drafting one — read the vertical, never guess it. On a creator-style spot, on a brief that asks for more hooks, or in a vertical the team has no pattern for, the cheap step is evidence. Pull 50–150 top-performing videos in the vertical over a bounded recent window sorted by engagement, transcribe the 10–15 most relevant, and tabulate one row each: views · the first spoken line, 0–3 s · when the product is first named · who is on camera. Then name the dominant pattern in ONE line, and say what the data said and what the pull cost. Search by TASK, not by vendor — the routed endpoints exist for "tiktok search videos", "instagram reels search by keyword", "tiktok video transcript" and the platform ad libraries, and a named competitor's live ads come from the ad library rather than from a guess. One measured pattern, dated and not universal: in agent and B2B-software verticals it runs stunt-or-claim → a specific number → the result → "here's how", with the product named LATE, as the answer (superdesigndev/treg make-ugc, read 2026-09-16).

A mined pattern shapes ORDER and FRAMING; it never replaces the client's line. The SSOT rule above does not bend for evidence: a hook drafted against the pattern is an addition, so it is a numbered cost line with "no" as the default, and the client picks. Draft more than will be used — 10 hooks, the client picks 3–5 — because picking is cheap and generating is not.

Size a talking hook to its take, not the take to the hook. A generated talking model fills every second it is given: too long and the delivery goes flat because there is nowhere to breathe, too short and words are dropped. Words ÷ 4, minus 1 second — 45–55 words ≈ a 12–14 s punchy hook, 60–70 ≈ 16–17 s, 70–80 ≈ 18–19 s and noticeably denser, 90+ splits into two clips because one take past ~20 s of continuous speech drifts. Join clauses with commas rather than full stops where no hold is wanted: the model reads a period as permission to stop for about half a second, which video-take-review's delivery row then reports as a HOLD. Confirm the script verbatim including slang, and warn the client that the on-screen captions follow the TRANSCRIPT — a transcriber normalises "gonna" to "going to".

Done when: the script file, the brand-kit table, the asset checklist and the decision log exist, every open question is a single numbered ask, and — where hooks were mined — the pattern table and its one-line verdict sit beside them.

3. Fix the global spec and the standing rules

9:16, safe zones (the declared targets' band — video-production SAFE-AREAS.md), the two-state grade, the lens, the music arc (the before cue → the wipe → the after cue → the closing bed), captions burned in, the venue's clip ceiling, the true-size framing rule. The seven standing rules — the product and its pieces DESIGNED never generated; text never generated; the explicit frame never framed; the wipe in every spot; B plates from the A keeper's frame 0; the fill SKU-locked and named on every shot; the easiest spot first and knowing what it proves — and the sign-off structure (the variable line's halves bracket the audio signature; the locked line; the CTA). A spot with no product, card or wipe on screen (a UGC testimonial) carries only the rules its script invokes — the scaffold's --spot-type ugc drops the rest. references/GLOBAL-SPEC.md.

The grade is named here, once, and it starts from the look library's cube for the genre (ads-clean for an ad) — also when the footage is the client's own: the operator picks it from a frame sheet of the project's frames (video-finish-qc/scripts/look_sheet.py --normalise auto) before the first delivered version, or hands the pick over and the agent picks on the sheet, provisionally. A hand-written level chain is a deviation that needs a reason; two of them stood for ten hours on a documentary spot before the operator asked (2026-09-15). The grade names its normalise step: a cube is the grade, never the normalise — flat sources (overcast light, full-range phone clips, a re-encode) are levelled and brought to a measured saturation per shot first (video-finish-qc/scripts/normalise_shots.py), and the sheet renders every cube on the normalised frame; on the cube alone an auction spot read pale beside a hand-normalised chain (2026-09-16). A sheet pick is provisional until one delivered spot has been watched in motion at full size. And the finish is part of the job: every render a project builder makes before video-finish-qc has run is PRE-FINISH, and says so.

Done when: the global spec section is written once for the campaign, names the look picked from the sheet and its normalise step, and every later spot cites it.

4. Build the asset bible, the kit and the beats — then the shot list

python3 ~/.claude/skills/video-production/scripts/script_diff.py --script prompts/<SPOT>-SCRIPT-client.txt prompts/CLIENT-NOTES-*.txt --beats prompts/<SPOT>-beats.json --elements ELEM-FILL-<colour>
python3 ~/.claude/skills/ad-spot-preprod/scripts/shotlist_scaffold.py --beats prompts/<SPOT>-beats.json --out prompts/r2v/<SPOT>-SHOTLIST.md --spot-type product|ugc --max-duration <read from the vendor> --sku "<SKU>"
python3 ~/.claude/skills/video-production/scripts/script_diff.py --script prompts/<SPOT>-SCRIPT-client.txt --beats prompts/<SPOT>-beats.json --shots prompts/r2v/<SPOT>-SHOTLIST.md --max-duration <the same cap>
  • The venue is chosen by a costed Phase-0 test on the HARDEST shots (a photoreal reference call, a lipsync line, the most exposed premise), cost per keeper as the unit — never by argument, never on the easiest spot.
  • The asset bible: characters (one entry per role, locked details, a count check against the script), products, A/B locations, recurring elements with their owners.
  • The pre-production core runs here, in its order (video-production/references/PREPRODUCTION-CORE.md): beats.json from the client's script FIRST, never from a prior cut; the generator's maximum duration READ before the shot list exists and written in its header; the script partitioned on CONTINUITY, never on shots — one generation per partition within the cap; with voice-over, the lines timed before any duration is fixed.
  • A format the team has not measured → tear down 3–5 real references BEFORE the shot list. The reference's STRUCTURE is the asset, never its pixels: scripts/reference_teardown.py --refs <a> <b> <c> --out review/teardown-<format> --runtime <s> measures every shot (length, framing by face size, motion, the words with --asr) and the rhythm across the set, and writes a timestamped beat grid at that rhythm; the content of each shot is then written by hand (or drafted by a multimodal lane, never its durations), what is kept and what is swapped goes into a table, and each slot takes ONE of our beats — the client's script stays the SSOT, and with a voice-over the VO's phrases cut and the grid only checks the rhythm: references/REFERENCE-TEARDOWN.md.
  • Then the per-spot shot list from the scaffold (--spot-type product for a product spot, ugc for one with no product, card or wipe): the method line, the cap, the kit table, the room geometry pin, the cast (closed after the first keeper), the PARTITIONS with the script lines each serves in full, the audio plan, the acceptance rows, the submission commands. Ids are stable. The next spot is the one that reuses the most accepted assets; the capper draws from the CURRENT accepted spots.
  • script_diff.py (core § 5) runs before the first prompt, both ways; in an ad campaign --elements also fails a piece-bearing beat without its element id — the one that changes later. references/SHOTLIST-CONTRACT.md.
  • Existing footage (the client supplies the picture): the same core with the generation stages declared not applicable, with the reason — the sources probed for their DISPLAY shape and transcribed whole first, the beats from the client's text or the brief's message lines, the shot list as a keeper list with a microphone-position column, the scaffold as --spot-type ugc with the gens table, the refs gate and the fill marked N/A in the header: references/INTAKE.md § Existing footage. Choosing the shots out of the client's pack is a phase with its own procedure — the frame-rate decision with its motion cost stated first, every clip broken down and read whole, the pack grouped into camera ROUTES (no route + subject twice), picture slots measured from the programme audio's phrases, candidate windows read at the delivery shape, relevance to the words before looks, a labelled proxy over the mix for the operator's pick: references/FOOTAGE-CURATION.md.
  • A UGC / creator-style spot (--spot-type ugc, phone-shot or generated — a person talking to the camera): the grammar is references/UGC-GRAMMAR.md — the intake axes (AOV band, funnel stage, platform, real or synthetic persona, phone-shot or generated), the hook core, the archetypes, the beats by duration with REAL proof composited at the demo beat, the delivery and caption rules, the variant matrix, the four artifacts; the script carries a disclosure column, and the compliance path is chosen at intake (references/RISKS.md § UGC compliance). A synthetic persona presents a demo, never a testimonial.

Done when: SCRIPT-DIFF PASS on the beats AND on the shot list (--shots), the generator's cap is in the header with its source, every section is filled, every partition row quotes the script lines it serves, and — for a new format — TEARDOWN.md holds the measured references, the written read, the kept/swapped table and the filled grid.

5. Write the risk register and the fallbacks

Filters (route around, never rewrite the joke), moderation on a word (the dub), lipsync (decided per line), transformation geometry, text in frame, packaging fidelity, likeness, platform copy risk (the fallback register built with the originals and parked), fill continuity, venue cost at the campaign's shot count. Open decisions numbered with an owner and a closing trigger. references/RISKS.md.

Done when: every risk names the mitigation that is already in the shot list, and the fallback register lists what each fallback costs.

6. Cost the rounds and hand over

python3 ~/.claude/skills/ad-spot-preprod/scripts/cost_plan.py --scenes prompts/r2v/<SPOT>-scenes.json --rate 2.5 --extra "stills:0.05" --extra "VO 954 chars:0.12"

Rounds by dependency (what runs in parallel, what waits for a pick), the spot total clean and with one re-roll per scene, stills and VO as their own lines — as numbered questions with the cost inline, one GO per round. The client-facing shot list ships as a designed PDF when asked. Then the hand-off: video-refs-continuity builds the references the first scene needs; video-gen-cost-gate submits.

A designed device gets a CREATIVE CHECKPOINT before the full build. When the spot's idea rests on something designed — type over footage, an animated opener, a card system — the first round is a short motion test of that device (a few seconds, one raster, the real footage and the real words) plus the typeface as an operator PICK from rendered options (the message line set in three to five candidates on the project's own frames), before polish, before the second raster, before any cut-down. A written standard ("brand-appropriate, licensed") does not replace the pick: on one job the operator's first look at the device in motion was the fourth full build, and every version to that point was rejected on the typeface and the device together (2026-09-15). A campaign whose kit is already accepted skips this.

A change to an accepted style is DERIVED from it, and the checkpoint reads the change. A new season, a second product in the same campaign, a client note on the look: the base is this subject's own accepted brief (designed-elements STYLE-BRIEF), never a catalogue style or another subject's brief.

  • Sort each wanted change by the brief's section: safe — the palette within its colour logic (§ 3), type sizes and roles (§ 4), the motion's tempo or how much moves (§ 5), the sound's density (§ 7), one native move added or swapped (§ 8); changes its character — materials and rendering (§ 2), the colour logic itself (§ 3), the camera grammar (§ 6), each with one line on what the style loses; not a change — the essence and what it is not (§ 1), the renderer (§ 10): that is a new brief, written from the subject.
  • Write a CHANGES table beside the brief, one row per change: section · base · now · why. The operator approves the table; the motion test is read against it.
  • At most five changes a round; more cannot be judged.
  • Render one style frame on the project's own frames beside the same frame in the accepted style, then the checkpoint's motion test (a few seconds, one raster, the real footage and words) with every change on screen at least once; name the second each change can be seen.

Done when: the cost lines are answered, the first round's references are listed by name, and the plan is saved in the project with its path in the ask.

Failure behavior

  • A script line quoted in a beat that is not in the client's text → FAIL; the beat is rewritten from the text, never the text from the beat.
  • A client note that contradicts the saved script → the newer client text wins; save it, diff, rebuild the beats, and say what changed.
  • A venue test that refuses the reference or the premise → the method changes (references for everything but the character; the line in post audio), the creative rulings do not.
  • A stale cost in a decision → re-cost before quoting; never cite an old figure as a reason.
  • A sentence or a partition split across rows with no recorded reason → one generation, or the reason written in the row (consistency not required across the split, or the action outruns the read cap).
  • A shot map taken from a prior cut's scene detection → discarded; the beats and partitions are rebuilt from the script.

Scripts

scriptdoes
video-production/scripts/script_diff.py --script --beats [--cover --cover-section] [--elements] [--shots --max-duration]quotes token-exact against the client's text (FAIL), uncovered sentences, unnamed fills, and the segmentation gate — a sentence or a partition split across rows (FAIL); --selftest
shotlist_scaffold.py --beats --out [--spot-type product|ugc] [--max-duration] [--sku]the per-spot shot list skeleton — the cap in its header, partition rows with the script lines in full
cost_plan.py --scenes [--rate] [--extra]rounds by dependency, the numbered cost lines, clean and worst case
reference_teardown.py --refs … --out [--runtime] [--card] [--asr small.en] [--dense 12] · --selftestthe measured structure of 3–5 real references in display geometry — shots (a cut = the coarse layout jumps AND the histogram jumps or the frame decorrelates hard), framing by the largest verified face, motion by phase correlation, the words with --asr — the rhythm across the set, a timestamped beat grid at that rhythm, TEARDOWN.md with the CONTENT column to write, one contact sheet per reference; --dense: a strip per shot (paged) of its frames at that rate between its cuts, for a motion-design reference; never overwrites

Cross-references

  • references/INTAKE.md · references/FOOTAGE-CURATION.md · references/REFERENCE-TEARDOWN.md · references/GLOBAL-SPEC.md · references/SHOTLIST-CONTRACT.md · references/RISKS.md.
  • video-edit-edl — the beats contract and the gate; video-refs-continuity — the reference set the plan names; video-gen-cost-gate — the cost line form; designed-elements — the kit's designed pieces; client-rounds — what happens when the client's notes come back.
  • Handoffs follow ~/.claude/skills/video-production/references/CHAIN.md.
  • Deeper context — the upstream projects behind the rules here: video-production/references/CONTEXT-MAP.md § Where the deeper context lives.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.