creator-content-ops — make your AI stack sound like YOU, across every channel
You are the example. If you run a page, a channel, a blog, you already have a voice. The problem is your AI doesn't. Ask it to "write like me" and it writes like the internet: ten drafts, ten different people, none of them you. Ask it to "gen a carousel" and the Thai fonts shatter, card 3 looks nothing like card 4, and every regen is a coin flip. That is not a model problem. It is a missing-system problem.
This kit is the system. Not a Canva brandbook with a color palette. A rule set your AI reads before it writes anything, plus the production layer that renders instead of gambles. Four formats, one spine:
- CONTENT.md — your voice, extracted once, loaded every session. The 3 filters that catch AI-smell before it ships. This is the heart. Everything else reads it.
- carousel/ — TikTok/IG carousels where AI gens the background and CODE renders the text. Thai never breaks, because it is rendered with a real font, not generated.
- shorts/ — YouTube Shorts / TikTok clip assembly from your raw material.
- covers/ — post covers, two paths: a code-composed neutral template (pixel-exact, free, every time) and a Codex ImageGen path (creative, one-offs, diagrams).
The split that makes it a factory, not a craft:
AI gens what needs imagination. Code composes what needs consistency. Backgrounds, art, hero shots → AI. Text, layout, export → code. Never let AI draw the words on the card. Render them.
When to use / not use
Use when a creator wants to set up an AI content stack, or fix these symptoms:
- output reads generic / "sounds like AI" no matter how you prompt it
- Thai (or any complex-script) text breaks when AI generates the image
- a carousel/series comes out inconsistent card-to-card
- every asset is hand-made one at a time in Canva and it doesn't scale
Do NOT use for: dev-repo setup (that's claude-md-setup — same bones, code target),
one-off single images with no voice/consistency need, or video editing craft. This kit
is about turning repeatable content into a governed pipeline.
Setup first — build YOUR CONTENT.md before anything else
Do not copy someone else's voice file. Run the interview in SETUP.md — it asks you
for your writing samples, your no-go words, your channels, and your tool auth, then
writes a CONTENT.md calibrated to you. Ten minutes. Every module below reads the file
it produces, so nothing works right until it exists.
1. Read SETUP.md → answer the interview, hand over 5-10 real samples 2. It writes CONTENT.md → your voice spine (3 filters + per-platform rows) 3. Wire the modules → point carousel/shorts/covers at your fonts + CONTENT.md 4. Produce → every format now reads CONTENT.md before it generates
The modules
| Module | What it does | AI or code | Deps |
|---|---|---|---|
CONTENT.md | voice spine + 3 AI-smell filters, read every session | — | none |
reference/carousel_factory.py | JSON → PIL render, Thai/Latin tofu-fix | code | Pillow, fonts |
reference/cover_neutral.py | pixel-exact cover from a template you rebrand | code | Pillow, fonts |
reference/codex_imagegen.md | creative backgrounds / diagrams via Codex CLI | AI | Codex CLI authed |
reference/shorts_pipeline.md | Shorts assembly from raw material | code + AI | ffmpeg, Pillow |
Principles
- You are the training data. The kit does not invent a voice. It extracts yours from real samples and enforces it. Garbage samples in, generic voice out.
- One spine, many formats. A carousel and a Short and a blog post all read the same CONTENT.md. That is why they sound like one person instead of four tools.
- The kit ships the method, not your moat. Copy the render tricks and the voice system. Your finished channel, your trained voice model, your private auth stay yours.
Portable across agents
Nothing here is Claude-only. CONTENT.md is plain markdown any agent reads; the render
engines are plain Python. Claude Code loads the voice file with @CONTENT.md; Cowork,
Cursor, and Amp each have their own load step (a context file, an AGENTS.md, a settings
path). The method is portable, the wiring per agent is one line.
