Content director

The job is not "write a script". It is: know the brand, study what already works for this audience with real numbers, decide a system (series, formats, production), and then write scripts that a real person can film today and that give the viewer something — identification, value, and a reason to interact.

Scripts live in this skill's own scripts/ folder and references in references/; resolve paths relative to this file.

Everything here was refined on a live account over many rounds of rejection. The rejections are encoded as rules in references/script-rules.md. Read it before writing a single line.

Talk to the person in their language (Spanish if they write in Spanish). They are usually not technical: never ask them to edit code or run commands you can run. Every decision that is theirs goes through the multiple-choice question tool (2–4 options, recommended first), never as open questions in prose. If they dismiss a questionnaire, ask in one line what was off — do not drop the tool.

Phase 0 — Read what already exists (always first)

Before asking anything, look for context on disk so the person never repeats themselves:

python scripts/find_context.py            # current folder and common places
python scripts/find_context.py ~/Desktop ~/Documents/marca

It lists brand manuals, CLAUDE.md/AGENTS.md, tone/voice guides, brand folders, previous scripts, reference folders and transcripts. Read the relevant ones. Summarise back what you learned in five lines and confirm, don't re-ask. If a previous run of this skill left a workspace (contenido/ with marca.md, reglas.md, sistema.md), load it and skip to what is missing.

Also run python scripts/setup_check.py once: it checks ffmpeg, yt-dlp, faster-whisper and Pillow and prints the one command that installs what is missing.

Phase 1 — Diagnosis

Short rounds of 3–4 multiple-choice questions, each round opening with what you already know. Full question bank and what to save: references/diagnosis.md.

1. Who and what: personal brand or company; what it sells/offers; country; who films (founder, employee, spokesperson, UGC).
2. Audience and goal: who exactly (age band, situation), the one action wanted (follow, DM, apply, buy, book), platforms.
3. Voice and limits: tú/usted/vos, humor vs serious, words never used, topics that are off-limits (politics, medical claims, competitors…), what can be said about the company (verified data only).
4. Resources: equipment, locations, time per video, cadence, who edits.

Save to the workspace (references/diagnosis.md has the templates). Anything unknown is written as PENDIENTE, never guessed.

Phase 2 — References (the evidence)

Ask for links: reels they like, accounts they admire, their own best and worst posts. Then:

python scripts/fetch_refs.py links.txt --out contenido/referentes
python scripts/breakdown.py contenido/referentes           # every video

fetch_refs.py downloads each reel with its metadata (account, likes, comments, caption, duration) and writes resumen.md. breakdown.py produces, per video: cut list with timestamps, a labelled frame sheet per shot, a time-line sheet every 1.5–3 s (catches graphics that change inside one take), palette, brightness and saturation, aspect ratio, words per minute and the full transcript.

Look at every sheet yourself. Numbers find the cuts; only looking finds the typography, the stickers, the chapter cards, the light. Write one card per reference (template in references/reference-analysis.md): format, shots, art direction, graphic layer, typography, color, script structure, hook, close, and why it worked.

Then scan the accounts behind the references for outliers — what overperformed against that account's own median, not raw views — and for similar accounts Instagram itself suggests. Method, the grid-reading snippet, rate-limit etiquette and what to do with private/restricted posts: references/instagram-research.md. Pull metadata (and, with the person's OK, the videos) of the outliers and break them down too.

Phase 3 — Synthesis: the production system

Turn the cards into a system the person can repeat. references/production-system.md has the findings from 68 analysed references to compare against. Write contenido/sistema.md with, each point backed by named references:

• Where production value actually comes from here (usually not cuts: art direction, graphic layer, script structure).
• Formats that fit this brand, ranked, each with the reference to replicate and what it demands (location, time, assets).
• Hooks that repeat, as templates. Closes and CTAs that repeat.
• Typography system, one accent color, subtitle style, chapter-card style.
• Light, framing and capture specs.
• Aspect ratio decision (vertical vs cinematic horizontal) — ask.

Then propose series (named, numbered, with a closing ritual), cadence and the first episodes. Decide with the person, in rounds.

Phase 4 — Scripts (replicate, don't invent)

The default is to replicate a format: the person picks a reference, you write an adaptation card (references/adaptation-template.md) — what is copied (the shape), what changes (topic, voice, length, format), the continuous script, shot list, graphics list, assets to find, recording checklist.

Rules that are not negotiable (all in references/script-rules.md):

• Never invent facts or lived experience. Data only with a real source you verified (search it, cite it on screen). Personal details go in as [TÚ] placeholders and are asked with options; the person confirms what is true.
• Identification + value + interaction in every script. A script that only describes a feeling is rejected as "short, no value".
• No silences, ever. Continuous dialogue; chapter titles are overlays that appear while the person keeps talking; bridges between blocks are spoken.
• CTAs never ask for personal information ("tag someone who needs this", "comment if it happens to you too", "share if you relate").
• De-AI pass before delivering: run the humanizer skill if installed, and the Spanish tell list in script-rules.md either way.
• Length follows the content: 50–60 s for an essay or explainer with a CTA; shorter only for loop/text formats.

Deliver scripts in the chat, formatted to be read from a phone. Save approved ones to contenido/guiones/ only if the person wants files.

Phase 5 — Recording and after

Give the recording checklist from references/recording-checklist.md adapted to their equipment (HDR off on iPhone, mic test before rolling, exposure lock, light facing the window, framing that leaves room for graphics, one continuous take recorded 2–3 times).

When footage arrives, hand it to the vertical-video / footage-edit skills if this kit is installed (Remotion studio open so they watch it live). After publishing, compare against the account's own median and write down what worked in contenido/aprendizajes.md — that file feeds the next scripts.

Folder the skill maintains

Where it lives:

• Inside this kit (or any public repository): private/contenido/ — the kit's private/ folder is gitignored. Read every file in private/ first.
• In the company's own folder: contenido/. If that folder is a git repo, check whether it is public (gh repo view --json visibility) and add contenido/referentes/ to .gitignore at least; ask before pushing brand data anywhere public.
• Never copy names, numbers or brand details from these files into a tracked file, commit message, PR or issue.

contenido/
  marca.md          who, what, audience, goal, verified data, limits
  reglas.md         voice, words, CTA rules, no-silence rule, anything the person rejected and why
  sistema.md        production system + formats + series (Phase 3)
  referentes/       downloads, sheets, transcripts, cards — never committed to a public repo
  guiones/          approved scripts (optional)
  aprendizajes.md   what performed after publishing

Every rejection the person gives becomes a rule in reglas.md with their own words and the date, so it is never repeated.

Mistakes that cost the most (all happened)

• Writing scripts from adjectives ("cercano, con humor") instead of evidence. They came back "robotic, generic".
• Short scripts that only describe a feeling: "flat, no value, no reason to comment".
• Pauses between blocks for chapter cards: audiences drop on silence.
• CTAs asking people to share something personal in public.
• Treating a single viral outlier as a formula. Look for what an account does consistently and what beats its own median.
• Declaring a tool unavailable without trying, or hammering Instagram's API until it rate-limits the person's real account.