Persona Account

---
name: persona-account
description: Run an ongoing AI creator account with ClipUGC, one persona, posting consistently over time. Use when the user wants to create an AI influencer or virtual influencer for Instagram or TikTok, asks for "the next post", wants a content plan / posting grid / content pillars for an AI creator account, wants the same person to stay recognisable across every post, asks what a fitness / beauty / faith / study / food / mom / tech influencer would plausibly post, or wants to grow and keep feeding an AI creator account. Covers persona definition, per-niche content pillars, the next-post loop, grid composition, and the picture-before-video credit gate.
argument-hint: "[niche, or 'next post']"
---

# Persona Account

You are running ONE AI creator account over time: an Instagram/TikTok persona whose posts must read as the same person living a plausible life. Cast once, then post forever, the face never changes, the setting/outfit/pillar rotate.

This is a different job from a one-off app promo video:

| | `ugc-director` | `persona-account` (this skill) |
|---|---|---|
| Unit of work | one app → one promo campaign | one person → an endless feed |
| Success test | does it convert | does the account read as a real human |
| Output | finished UGC video (clip + app footage + hook) | post (look, usually + clip; occasionally a promo post) |
| Repeats | 2–3 hook variants of the same clip | a new post every time, never the same beat twice |

Use `ugc-director` when the user says "make a video for my app". Use this skill when they say "create an AI influencer", "give me the next post", or "grow my account".

## Reference files (read the ones you need)

- `references/niches.md`: persona profiles + content pillars per niche (fitness, beauty/GRWM, faith, study/productivity, food, mom-life, tech), with ready `--description` strings and what each persona would NOT post.
- `references/grid.md`: what a believable 6/9/12-post grid looks like, rotation matrices, sequencing rules, cadence, caption voice, and a grid audit checklist.

Borrow, never restate:

- **Command syntax, credit costs, prerequisites, the `clipugc/` workspace and `influencer.json`** → the `clipugc` skill.
- **Reaction archetypes (the 12), clip `--prompt` text, look `--scene` text, hook formulas** → the `ugc-director` skill and its `references/formats.md`, `references/prompts.md`, `references/hooks.md`. Copy clip prompts VERBATIM; do not write your own reaction prompts here.
- **The emotional palette for an ongoing grid** → ugc-director's `references/reactions.md`, ~30 silent reactions across 9 families (affection, longing, sulk, playful, sass, anxious, cozy, warmth, shock). This is the file this skill leans on most: the 12 promo archetypes are shock-and-curiosity shaped, and a grid built only from those reads as a bot. Real creator accounts run mostly on the soft families with a jaw-drop as punctuation.

The silent format is the same one `ugc-director` enforces. Its directing rule 1, verbatim:

> **Never prompt talking.** All clip prompts must include "lips closed / no talking", lip-sync is the AI tell. Text overlay does the talking.

On a persona account the caption (or the burned-in overlay on a promo post) is what does the talking. No exceptions, a persona account posts more clips than a promo campaign does, so one talking clip poisons a whole feed.

## Cost discipline: the approval gate (read before spending anything)

A picture is 2 credits. A clip is 7 (5s) or 13 (10s). A finished video (app recording + hook on a clip) is 1. So the cheap thing is the face and the expensive thing is the motion.

**The gate: look first, then ask, then spend.**

1. Generate the look (2 credits) and download it into `pictures/`.
2. Show the user the image path, plus what the post is (pillar, setting, caption direction) and what the clip will cost.
3. **Wait for an explicit yes.** Never run `videos create` in the same turn as `images generate`.
4. Only on approval, spend the 7 (or 13) credits on the clip.

A rejected look costs 2 credits. A rejected clip costs 9. **If the user does not like the character at all, no video credits are ever burned**, that is the entire point of the gate.

When a look is rejected, diagnose before re-spending:

| What's wrong | Fix | Cost |
|---|---|---|
| Setting, outfit, framing, mood | New `--scene` on the same character | 2 |
| This one look is close but off | `images variation <lookId> --scene "…" --count 2` | 2 each |
| The *person* is wrong (age, vibe, gender, face) | The casting is wrong, recast with a new `--description` (see `niches.md` and ugc-director's `casting.md`). Do NOT keep buying looks off a character the user dislikes | 2 (new character) |

Per-post maths:

| Post type | Commands | Credits |
|---|---|---|
| Photo post | `images generate` | 2 |
| Video post (5s) | `images generate` + `videos create --duration 5` | 9 |
| Video post (10s) | `images generate` + `videos create --duration 10` | 15 |
| Promo post (video + the user's app recording) | + `videos merge` | 10 |
| Scene-staged clip (`videos create --scene`, stages then animates) | replaces the two-step | 9, no separate look |
| Look the user rejected | `images generate` only | 2 |

**Do not use the scene-staged shortcut on a persona account.** It is one credit cheaper than look-then-clip, but it stages the new setting inside the video job, so there is no still to approve, the gate disappears and a bad setting costs 9 instead of 2. Take the two-step every time.

These are the current published values. **Read live costs with `clipugc credits` before quoting a budget**, costs are duration-aware and server-authoritative; never hard-code them into a plan you show the user.

## Step 1: Define the persona (once, then obey it)

Do this the first time the user says "create an AI influencer for <niche>". Ask only for what's missing; propose defaults from `niches.md` rather than interrogating the user.

| Field | What to settle | Source |
|---|---|---|
| `niche` | fitness, beauty, faith, study, food, mom-life, tech, … | user |
| `handle` | account name/handle the posts are for | user (optional) |
| `age_look` | age band, build, hair, one facial signature, authenticity anchor | `niches.md` → ready `--description` |
| `aesthetic` | colour/light palette the whole feed obeys (e.g. "warm neutrals, morning window light, film grain") | `niches.md` |
| `settings` | 4–6 recurring places this person actually lives in, the rotation pool | `niches.md` + ugc-director `prompts.md` scenes |
| `wardrobe` | 3–5 outfits the person re-wears, matched to the settings and the season | `niches.md` |
| `voice` | caption tone + energy (e.g. "low-hype, lowercase, first person, self-deprecating") | user/`niches.md` |
| `pillars` | 3–5 recurring post types | `niches.md` |
| `posts_about` / `never_posts` | the boundary that keeps the account plausible | `niches.md` |
| `visibility` | `--private` if this persona is the user's alone; public/discoverable is the CLI default | user |

**Cast for attraction here, not for a payer persona.** An account grows on whether people want to follow this person, so `age_look` should say it outright, "very pretty", "striking", "conventionally attractive", unless the user asks for something plainer. Pair it with the realism anchor or it reads synthetic: "natural skin texture", "not airbrushed", "reads as a real person". Makeup is directable ("barely-there makeup", "soft glam"); nothing forces a bare face. This is the one place `persona-account` deliberately parts from ugc-director's "cast the paying user, not the prettiest creator", that rule optimises app conversion, this account optimises follows.

**Start from a known-good casting rather than writing one from scratch.** ugc-director's
`references/casting.md` has ready-made, proven descriptions (blonde Scandinavian, sun-kissed
Californian, Mediterranean soft-glam, polished Korean) plus the four-part formula behind
them, concrete features, an explicit attractiveness claim, directed makeup AND a named
lighting setup, then the realism anchor. Adapt one to the niche instead of inventing a
description; a vague cast ("gen-z woman, brown hair, friendly smile") is the single most
common reason a persona's first look comes out looking generic. Cast adults only, and keep
the stated age unambiguously adult.

Then cast, once, and never again for this account:

```bash
# 2 credits, first look included. That first look becomes the BASE IMAGE: the identity
# anchor every later picture is generated from. Make it on-brand.
clipugc characters create --description "<persona description>" \
  --scene "<the persona's primary setting, from niches.md / prompts.md>" --wait --json
clipugc characters rename <characterId> --name "<Persona Name>"
```

Create `clipugc/influencers/<id>-<kebab-name>/`, write `influencer.json` with the `persona` block below, download the base look into `pictures/`, and record its id as `persona.base_image_id`.

**Mid-conversation persona edits are persona edits.** When the user says "she's a fitness influencer", "keep roughly the same outfit", "less makeup from now on", write it into the `persona` block immediately. Every later post reads that block first, a preference stated once must not need restating.

## Step 2: The next-post loop (the core of this skill)

When the user says "next post" (or "another one", "post for tomorrow"):

1. **Read the manifest.** `clipugc/influencers/*/influencer.json`, the `persona` block and the `posts` array. If the manifest marks anything pending/processing, or it looks stale, reconcile with the server first per the `clipugc` skill's resume flow (`characters show`, `images list --character <id>`, `videos list --character <id>`).
2. **Pick the pillar.** From `persona.pillars`, one NOT used in the last post; prefer the pillar least used across the whole `posts` array.
3. **Pick the setting and framing.** From `persona.settings`, never the setting used in the previous post, and prefer one unused in the last two (reuse it sooner only with a clearly different light or time of day). Rotate framing too (arm's-length selfie → mirror → mid-task glance-up → photo-only still).
4. **Pick the reaction archetype** from ugc-director `formats.md`, not the archetype used in either of the last two posts.
5. **Compose the look `--scene`**: persona `aesthetic` + the chosen setting + a wardrobe item + 3–5 authenticity keywords from `prompts.md`. **Never re-describe the person**, the base image carries the face; re-describing it causes drift.
   ```bash
   clipugc images generate --character <characterId> --scene "<setting + wardrobe + authenticity keywords>" --wait --json
   clipugc images download <lookId> -o clipugc/influencers/<id>-<name>/pictures/<lookId>-<short-desc>.png
   ```
   Record the look id + the exact `--scene` in the manifest's `pictures` array as soon as the command returns (the `clipugc` skill's record-ids-immediately rule).
6. **STOP. Show the look and the plan, and wait for the explicit yes** (see the approval gate above). Present: the image path, the pillar, the caption draft, the archetype, and the clip cost.
7. **On approval, and only then**, buy the motion, reaction prompt VERBATIM from `reactions.md` (or `prompts.md` for the promo archetypes). Pick a family the grid has not used recently; rotating feeling is what stops a grid looking generated:
   ```bash
   clipugc videos create --image <lookId> --prompt "<reaction prompt from reactions.md / prompts.md>" --duration 5 --wait --json
   clipugc videos download <videoId> -o clipugc/influencers/<id>-<name>/clips/<videoId>-i2v-5s.mp4
   ```
8. **Write the caption** in `persona.voice` (see `grid.md` for caption rules), and append the post to the manifest's `posts` array.
9. **Add the app recording only if this is a promo post**: i.e. the user has an app screen recording and wants this post to sell. Then it is a `ugc-director` job for the hook and the anatomy; the `videos merge` command and the finished video id space live in the `clipugc` skill (`merged_video_id` ≠ clip id).

Deliver ONE post per request. Do not batch a week of posts unless the user explicitly asks for a batch, and if they do, still gate every clip behind the look approval, look-by-look.

## Step 3: Continuity rules (what makes it one person)

1. **The base image never moves.** The character's first completed picture is the identity anchor; every later picture is an edit of it. Both `images generate --character <id>` and `images variation <lookId>` keep the face, choose by intent (a new look of the person vs. a remix of one specific look), not out of identity fear.
2. **Never recast mid-account.** A new character is a new person and breaks the illusion. Recast only when the user rejects the persona itself, and then start the account over.
3. **Rotate settings; never repeat a setting back-to-back.**
4. **Wardrobe must be plausible for the same life.** A real creator re-wears clothes: repeat outfits across the feed, but keep a single day/season coherent, no snow parka one post and a tank top the next unless time has passed in the story.
5. **Vary the reaction.** Same expression twice in a row reads as a template, not a person.
6. **One emotional register per persona.** A calm wellness persona does not suddenly do Speed-Shock double-takes.
7. **Respect `never_posts`.** The boundary is what makes the account believable; breaking it for one good idea costs more than the idea is worth.
8. **A failed generation is stochastic**: `retry` once before rewriting the prompt.

## Manifest: extend `influencer.json`, do not invent a new file

Same file, same folder, same rules as the `clipugc` skill's "Project workspace" section (record ids immediately, update after every step). Add two keys; keep `pictures` / `clips` / `finished` exactly as they are, they stay the id + prompt ledger, and `posts` only indexes into them.

```json
{
  "id": 12,
  "name": "Isabella Romero",
  "description": "warm approachable woman in her late 30s, athletic but realistic build …",
  "created_at": "2026-08-17",
  "visibility": "public",
  "persona": {
    "niche": "fitness",
    "handle": "@izzy.moves",
    "base_image_id": 87,
    "age_look": "late 30s, athletic-realistic build, freckles, practical ponytail",
    "aesthetic": "warm neutrals, morning window light, slight grain",
    "settings": ["home gym corner", "kitchen counter", "parked car", "bathroom mirror"],
    "wardrobe": ["grey workout top + black leggings", "oversized hoodie", "denim jacket"],
    "voice": "low-hype, lowercase, first person, honest about hard days",
    "pillars": ["morning routine", "progress check-in", "form tip", "meal of the day", "rest-day honesty"],
    "posts_about": ["consistency over intensity", "training around a full-time job"],
    "never_posts": ["extreme diets", "medical advice", "supplement miracle claims"]
  },
  "pictures": [
    { "id": 87, "prompt": "kitchen counter, early morning window light…", "file": "pictures/87-kitchen-base.png", "status": "completed" },
    { "id": 88, "prompt": "home gym corner, morning light…", "file": "pictures/88-home-gym.png", "status": "completed" }
  ],
  "clips": [
    { "id": 92, "mode": "i2v", "duration": 5, "prompt": "…approving closed-mouth half-smile, nods along slowly three times… No talking.", "source_image_id": 88, "file": "clips/92-i2v-5s.mp4", "status": "completed", "merged": false }
  ],
  "finished": [],
  "posts": [
    {
      "n": 4,
      "date": "2026-08-17",
      "pillar": "progress check-in",
      "setting": "home gym corner",
      "framing": "arm's-length selfie",
      "archetype": 9,
      "picture_id": 88,
      "clip_id": 92,
      "merged_video_id": null,
      "caption": "week 6. still not shredded. still going.",
      "status": "posted"
    }
  ]
}
```

`posts[].merged_video_id` stays `null` for organic posts and carries the finished video id for promo posts (the same id recorded in `finished[]`). Prompts live in `pictures`/`clips` only, never duplicate them into `posts`.

**Resuming an account**: read `influencer.json` first, reconcile with the server, and derive the next post from `posts` (last pillar, last setting, last archetype, post count). If a manifest is missing but the character exists on the server, rebuild the folder + manifest from `characters show` / `images list` / `videos list`, then infer a `persona` block from the existing looks and confirm it with the user before posting.

## Growing the account

`references/grid.md` has the blueprints. The short version:

- A believable grid mixes pillars, settings and framings, not one pose repeated in nine outfits.
- Establish the person in the first 3 posts (face-forward, niche legible), then rotate.
- Photo posts are legitimate posts and cost 2 credits; a feed of nothing but video reads as a content farm.
- Promo posts are at most 1 in 5, and never in the first 3.
- Budget a 9-post grid (6 video + 3 photo) at 2 (cast, first look included) + 8×2 (further looks) + 6×7 (clips) = **60 credits**. Confirm against `clipugc credits` before promising a number.

## When the account also sells an app

A promo post is a `ugc-director` deliverable dropped into this feed: its hook, archetype and 15–30s anatomy come from that skill's references, and the `videos merge` / finished video id handling from the `clipugc` skill. Keep the persona intact, the same face, the same voice, an app it would plausibly use. If more than one post in five is a promo post, say so and push back; a feed that only sells stops reading as a person.