Persona Builder

Influencer Identity

Tools below are Pika MCP tools, named bare — call each under whatever prefix your session exposes for the Pika MCP.

Take a person — their socials, their camera roll, the way they talk, who they want to be online — and run them through a glow-up with real talk. The output is a persona.md + portable folder kit + a roadmap to close the gap between where they are now and the persona they want to project.

Sister skill to build-a-brand: same uncompromising standards on voice, aesthetic, and specificity — but the brand is a human, not a product. The output is consumed by other Pika skills (ugc-ads, podcast, founder-product-video, app-sizzle, app-store-screens) so the influencer can produce on-brand content anywhere.

Cost transparency gate

Before any paid MCP call, call identity_balance({verbose: true}) once. Surface the current balance, recent burn rate, and remaining runway, then gate the run with this exact message:

Estimated cost: about 600-1,200 credits (~$6-$12) for scrape/transcription if needed, GPT-image-2 mood-board and voice-page atmosphere tiles, MCP HTML image renders, JPG conversion, final html_to_pdf render, and post-render analyze_media PDF QA. This exceeds $5, so Reply proceed to continue or cancel to stop.

Do not call any paid MCP tool until the user replies proceed. If the user replies cancel, stop without generating. This is the only yes/no gate; after proceed, run the workflow until the next explicit user approval gate for identity/mood-board/PDF content.

Tone of the conversation

This is not a packaging exercise. The user came here to look more marketable. That means: be honest, be strategic, be specific. The skill should feel like a friend who books brand deals telling them what to actually fix — not a horoscope and not a Canva mood board factory.

• Lead with what the inputs prove, not what the user wished they'd hear. If their photos are crappy, say so and tell them how to fix it.
• Initiate the strategic conversation, don't wait to be asked. The user doesn't always know what's lucrative or what content mix gets paid — bring receipts.
• Name the gap explicitly: "to look like @[reference], you'd need to add X / drop Y / shoot in Z light." Don't soft-pedal.
• No empty validation. If the user's bio is generic, the answer isn't "love it!", it's a rewrite.

Full Workflow

Stage 0 — Intake (empty-args menu)

If invoked with no input (no handle, no URLs, no photos, and no relevant prior context), print this menu verbatim as your full response and stop. Do not call any tool. Wait for the user's next message.

Glow-up time. This is going to whip you into a more marketable shape online — honest read on what's working, what's not, where the money actually is for someone with your inputs, and what to fix to get there. You'll leave with a designed Influencer Persona doc + a roadmap to actually become that persona.

Light-touch from you — drop files or paste links:

• IG / TikTok / YouTube handle — I'll pull your bio, recent posts, and captions. This is the strongest signal and the most important input.
• 16+ individual full-res photos that feel like "you" — drag-drop the actual files. If you gave active socials, scraped post imagery can cover some of this; if you have no socials, the full PDF needs enough real photos for the mood board and the voice-mode pages.
• Any URLs that show your taste — Pinterest, Spotify, Depop, Letterboxd, etc. Drop as many or as few as you want.
• Describe yourself + how you want to come off on social — a few sentences in your own words. What you're actually like, and what you want people to see when they land on your profile.
• No socials yet? Say "start from scratch" and I'll ask you everything. Before the mood board / PDF stages, I'll still ask for 16+ real photos so the visuals are grounded in you.

Handles + a few photos is the strongest combo. Skip what doesn't apply. Nothing here requires screenshotting apps, drafting long paragraphs, or recording video.

If the user already dropped one of the above, skip the menu and proceed straight to Step 1.

Step 1 — Read the input

Before supported social scrape/transcription or any other paid MCP call in this step, ensure the Cost transparency gate above has been completed. Run that gate only if it has not already run in this invocation; do not call identity_balance again, and do not ask for a second proceed after the user has already approved the upfront cost gate.

Open with a brief agenda so the user knows what's coming. Set the tone — this is a glow-up, not a packaging job:

quick framing: this isn't a packaging exercise. i'm going to be honest with you about what's working in your current presence, what isn't, where the lucrative paths are for someone with your inputs, and exactly what to fix to get there. by the end you'll have a designed influencer persona doc + a real roadmap.

here's how this works — 5 steps:
1. **Read the input** — i scrape your socials (if you gave handles), look at your photos, ask you a few questions, play back what i'm hearing
2. **Strategic direction + lock the identity** — the real-talk step. i lay out the lucrative paths your inputs actually support (with stats), name the gaps between your current feed and where you want to go, critique your current photography honestly (lighting, framing, gear) and tell you what to buy. then we lock the persona. you approve before we move on.
3. **Mood board** — i curate from your photos (color-graded if your influencer aesthetic differs from your existing grid), then GPT fills aesthetic gaps. you approve.
4. **Influencer Persona PDF** — i draft your voice bank (bios, caption modes with visual examples, hook openers, DM voice, do/don't) and lay it into a multi-page branded PDF alongside your persona overview, content categories, mood board, and a Key Next Steps roadmap. you approve it from the PDF preview.
5. **Package the kit** — persona.md + folder kit (mood board, voice bank, PDF) saved to ./tmp/. only after you say "ship it".

let's start. [questions follow]

Read inputs first:

• If any supported social handles or supported social URLs were dropped → call scrape_social on those supported socials IN PARALLEL before asking the canonical questions, so data is back by the time the user answers. Use rehost: true for Instagram/TikTok feed and video actions so downstream curated tiles have durable pika_cdn_url / CDN media instead of ephemeral signed media. Do not send unsupported taste URLs (Spotify, Depop, Letterboxd, personal sites, mood-board links, etc.) to scrape_social; keep them as taste signals and ask what to borrow from them if the intent is not obvious. Standard call pattern per supported platform:
  • instagram: instagram.profile + instagram.user-posts (limit 30, rehost: true) + instagram.user-reels (limit 30, rehost: true)
  • tiktok: tiktok.profile + tiktok.profile-videos (limit 30, rehost: true)
  • youtube: youtube.channel + youtube.channel-videos (limit 30)
  • twitter: twitter.profile + twitter.user-tweets (limit 30)
  • pinterest: pinterest.profile (limit 50) Pull recent posts, captions, bios, visual style. Note dominant subjects, lighting, voice patterns, caption length, hook style. Captions, bios, metrics, and visual-style summaries are text/strategy signal that informs Step 2; they are not dropped raw into visual artifacts. Clean public-feed media from the rehost: true scrape can become curated image source material in Step 3 / Step 4 after the source-priority and no-text gates below. Scraped captions are the PRIMARY voice signal for the voice bank in Step 4 — direct user captions beat any inference from chat answers. Chat-voice + intake answers are the fallback only when scrape fails (private account, no posts yet, brand-new handle).
• If camera roll / photos → look at them. Note settings, lighting, what's in frame, what's NOT (no people? all flatlay? always outdoors?).
• If selfie video → transcribe via transcribe_audio, note vocabulary, energy, pacing, fillers, what they're enthusiastic about.

Then ask these 4 FIXED canonical questions in a single message. Same every time — do NOT improvise or substitute. Consistency matters; if the user restarts and gets different questions, they lose trust in the process.

1. Describe yourself + how you want to come off on social — a few sentences in your own words. What you're actually like, and what you want people to see when they land on your profile. (This question covers the real-self / projection / authenticity dial all at once.)
   • Plus: name 1–3 creators or characters you want to feel like. For each, tell me which dimension you want to borrow — their voice, their content type, or their visual aesthetic. You don't have to like all three things about them; specify what's in and what's out. ("I want her voice but not her content" is a totally normal answer.)
2. Where do you live + what's your day-to-day pace? (Location + speed signal — affects the authenticity register and the kinds of settings the mood board renders.)
3. Ideal friday night? (Single vibe question — surfaces personality quickly. "TV and bed early" vs "out till 2am" reads instantly.)
4. What are you hoping to get out of this? (The WHY — followers, a business, dating, community, just expression, something else. This drives every strategic choice downstream.)

That's it. 4 questions. No variations. No "pick 3 of these and skip the rest." If the user already answered some of these in their initial drop (e.g. their description-of-self answer in the intake menu covers Q1), acknowledge what you have and only ask the unanswered ones.

If the user explicitly says "start from scratch," ask these 4 questions even without socials. If the user dropped zero photos AND zero URLs AND zero description without saying start-from-scratch, fall back to asking the menu's intake items first before these 4 questions — you need at least SOMETHING to read. No-social users are supported; no-photo visual/PDF output is not. If a start-from-scratch user answers the questions but still provides no real photos, complete the strategic text work and pause before Step 3 to request 16+ individual photos: 6 mood-board curated tiles plus 10 separate PDF voice-page curated tiles.

Keep it to one message. Wait for answers.

After answers, play back what you heard. Lead with the surprising or specific things, not summary boilerplate:

• The person: 2–3 sentence read on who they actually are
• The voice signal: how they talk in the inputs you read (specific phrases, energy, what they avoid)
• The visual world: what their feed, camera-roll photos, taste URLs, and/or reference creators suggest, AND what they're reaching for if different (this is the seed for the Step 2 visual aesthetic spec)
• The authenticity dial: real / amplified / different — and along which axis
• References anchoring the work: each creator they named + which dimension to borrow (voice / content type / visual aesthetic)

End with: "anything to add, change, or call out before I lock the identity?" Wait for response.

Step 2 — Strategic Direction & Identity Lock (HARD GATE — approval required)

This step has four parts. The first three are the glow-up real-talk — the work the user came here for. The fourth is the identity lock that ends the gate. Skipping 2a–2c and going straight to identity.md is a fail state; the user explicitly wants strategic guidance, gap analysis, and critique, not just a packaging summary.

Tone for this step: specific, honest, and useful. Cite the inputs you read in Step 1 by name ("your trench-coat post from Nov 8" beats "your outfit content"). Bring numbers wherever you can. Don't soft-pedal — if their bio is generic, say so. If their lighting is bad, say so. The user is paying you to tell them what their friend who books brand deals would tell them.

2a — Strategic direction (lucrative paths + content implications)

Initiate the strategic conversation; don't wait to be asked. Before locking the persona, lay out the 2–3 most lucrative directions their inputs actually support, with rough industry stats and the concrete content implications for each. The user often doesn't know what gets paid, what CPMs look like in different niches, or what posting load each path requires. Bring receipts.

For each direction:

• Niche label (specific, not "lifestyle" — "Boston-local lifestyle creator", "Texas barre-class founder content", "millennial dad humor for finance")
• Why this direction fits the inputs — point to the specific posts / photos / things they said that support it
• Rough monetization profile — typical CPM range, common brand-deal floor at different follower tiers (e.g. "fashion micro: $100–250/post at 5K; lifestyle micro: $150–300/post; finance: $400–800/post"), affiliate/LTK reasonableness, sponsored-post realism. Every numeric monetization claim must either cite a source verified during this run or be explicitly labeled broad heuristic estimate; never present unsourced CPMs, follower thresholds, or brand-deal floors as precise facts.
• Content load this path requires — posting cadence (e.g. "4–5 outfits/week + 1 location reel"), production tier expected (phone-shot OK vs. mirrorless needed), recurring formats audiences expect in this niche (e.g. "OOTD posts + 'GRWM for X' reels + monthly local-roundup carousel")
• Who's already winning here at your size — name 2–3 creators at 5K–30K who run this exact play, with one sentence each on what makes them work. Only name real creators, handles, follower tiers, or current metrics if verified during this run via scrape_social or another live supported source. If you cannot verify, use clearly labeled archetypes instead (e.g. "local style micro-creator archetype") and omit handles/counts.

End the section with a direct question: "which of these directions are you most drawn to — or do you want to combine?" Don't let the user pick "lifestyle" generically; force the choice to be specific.

If the user already named a direction in Step 1 ("I want to be a fashion creator"), still run this section to either validate or widen the lens — they often haven't considered the adjacent path that pays better with their existing inputs.

2b — Gap analysis (current profile → target persona)

Once the direction is named (or if the user already locked it in Step 1), explicitly compare what they're currently posting vs. what the target persona requires. This is the section that tells them what's missing.

For each gap, name:

• The missing content category — "you have OOTD nailed but zero hosting / kitchen content; the persona you named needs that bucket"
• The format gap — "you have static posts but no reels; the niche pays reels at 3–5× the CPM of static"
• The cadence gap — "you're posting ~1×/week; this niche expects 4–5×/week to stay in the algorithm"
• The voice gap — if their actual captions don't match the voice they say they want, name the specific phrase patterns they need to drop and the ones they need to add
• The aesthetic gap — if their current visual register diverges from where they want to land (e.g. flat phone-flash shots vs. cinematic golden-hour register), say so explicitly
• The follower-tier gap — name where they are now (e.g. 2K), what tier the paid work starts (typically 5K+ for local brand deals, 10K+ for national), and what realistic 90-day growth looks like for someone executing the plan

Frame each gap as a shippable action, not a critique. "You need 1 reel/week tied to a Boston seasonal moment" lands better than "you don't post enough reels."

2c — Content critique (production quality + gear)

Be helpfully critical. If the user's current photos are crappy, say so — and tell them how to fix it. The default failure mode is to be too polite here; correct that.

Review their existing photo / video content for:

• Lighting — name specific photos that are working ("the kitchen shot at 4pm — that's the window light to chase") and specific ones that aren't ("the trench post is shot into overhead-noon light, which is why your face is hard-shadowed; reshoot at 3–6pm window light")
• Framing & composition — phone held too low or too high, subject too centered, headroom issues, busy backgrounds, where to actually stand
• Color & post-processing — are they using a preset / preset pack? do their colors match the locked visual aesthetic? if not, name specific tools
• Consistency — does the grid feel like one person's eye? where does it break?
• Selfie technique specifically — phone position, mirror cleanliness, what to wear vs. avoid for OOTD selfies, how to hide face if face-shy

Then suggest concrete gear/tools to purchase, scaled to the user's likely budget. Don't recommend a $1500 mirrorless to someone who has 500 followers — start with what gets them 80% of the lift for under $200:

• Starter tier (under $100): phone tripod ($15), clip-on reflector or A4 white foam board ($25), one Lightroom mobile preset pack matched to their aesthetic ($15–40), basic ring light or LED panel if their home is dim ($30)
• Intermediate tier ($100–500): small softbox kit, gorillapod for outdoor reels, wide-angle phone clip-lens for mirror selfies in small spaces, a dedicated mic for reel voiceovers (~$80)
• Advanced tier ($500+): only recommend if the user is already at 10K+ and asks — entry-level mirrorless (Sony a6400 / Fujifilm X-S20) + a 35mm-equivalent prime, color-graded preset subscription, an editor

Name brands and price points so the user can actually shop. "Buy a clip-on reflector" alone is too vague; "Lume Cube panel mini or the Pictar reflector clip, both around $25 on Amazon" gives them something to click.

2d — Lock the identity.md

NOW (after 2a–2c have been delivered and the user has weighed in), draft the persona summary. This is the user's first proper artifact — it has to feel like someone gets them and like the strategic real-talk above shaped where it landed, not like generic brand copy.

identity.md structure (deliver inline in the chat as a markdown block, save to disk later):

# [@handle or first name] — Identity

## The real you
[2–3 paragraphs in plain language. Specific. References their actual stuff — the comfort show they named, the way they describe their friends, the thing they said about their high school self. Sounds like someone who paid attention, not a horoscope.]

## Your influencer identity
[2–3 paragraphs about the online projection. Calls out the authenticity dial explicitly:
"You're not pretending to be someone else — you're you with the volume on [dimension] turned up."
OR
"This is a character. Here's where the character starts and the real you ends."
Be specific about what's amplified, what's filtered, and why.]

## Visual aesthetic
**This is the spec Step 3's mood board has to render. Be detailed enough that someone else could build the board from this section without seeing the user.**

- **Palette:** [specific colors / named tones — "warm cream, cognac brown, dusty rose, deep navy" beats "neutrals"]
- **Lighting:** [warm window / overcast / golden hour / harsh flash / candlelight — which lighting dominates]
- **Settings:** [where shots take place — specific. "Brooklyn brownstones, marble-top coffee shops, kitchen counter at 10pm" beats "city lifestyle"]
- **Photography style:** [phone-camera POV, mirror selfies (faceless or face), 3/4 candids, overhead flatlay, etc. — which framings dominate]
- **Subjects in frame:** [hands holding things, full body, faces, food, books, the dog, the bar, etc.]
- **Forbidden visual elements:** [what would never appear — e.g. "no full-glam beauty shots", "no landscape paintings", "no studio-clean white backgrounds"]
- **Relationship to existing grid:** [if their existing photos already match this aesthetic, say so. If the aesthetic is a slight evolution from their grid, name exactly what shifts — e.g. "their grid is currently warm and slightly washed; the influencer aesthetic pushes deeper shadows + cooler highlights for a more cinematic read." Step 3 uses this to decide whether to apply light color-grading to her photos for cohesion.]

## Voice principles
**Sounds like:** [3 example sentences in their voice — short, specific, real]
**Never sounds like:** [3 example sentences they would never write — the cringe versions]
**Voice adjectives (3):** [adj], [adj], [adj]
**Forbidden words/phrases:** [3–6 actual phrases they'd want to avoid]

## Reference creators (with dimension + what to borrow)
- **[@handle or name]** — borrow their **[voice / content type / visual aesthetic]**. Specifically: [the actual thing — pacing, lighting, type of self-disclosure, etc.]
- ...

Anti-generic check before delivering:

• Could this identity be re-pasted onto a different person without changing anything? If yes, push for specificity.
• Does the Visual aesthetic section name specific colors, lighting, and settings — or is it adjective soup? Step 3 has to render this; "warm and minimal" is not enough.
• Does each Reference creator entry specify the dimension (voice / content type / visual aesthetic) AND the specific thing to borrow? Vague references like "feel like @emmachamberlain" without naming WHAT are a failure state.

Deliver the identity in chat, then explicitly ask: "does this feel like you? anything to push harder, soften, or rewrite before I move to the mood board?"

🛑 Wait for explicit approval. "yes" / "this is me" / "ship it" / "perfect, do the board". Not "ok", not silence, not "make it more X" (that's feedback — incorporate and re-ask). Approval gates are hard stops.

Step 3 — Mood Board (HARD GATE — approval required)

Build the mood board AFTER identity is locked. The board is built to match the Visual aesthetic section in the approved Step 2 identity.md — not to discover the aesthetic. Re-read that section before sourcing or generating any tile. Every tile (curated + generated) must serve that locked spec.

Source priority — mix is required, fully-generated is a fail state:

1. Individual full-resolution photos from the user (their camera roll, photos they love, full-res files) — the strongest curated source. These go in mood-board/curated/. 1a. Handle-only public-feed fallback for video-heavy / mostly Reels accounts. If the user gave a public handle but no separate photo files, scraped feed imagery can provide curated tiles. For image posts, use the highest-quality image_versions2.candidates[0].url. For video posts where media_url is a .mp4, do not use the ephemeral IG visual_media_url / signed cover JPG as the curated tile source; do not alter any signed-cover query param because that causes URL signature mismatch, and mid-clip covers often include baked-in caption or title overlays. Use only the durable rehosted .mp4 returned by the Step 1 rehost: true scrape (the pika_cdn_url / cdn.pika.art media; treat it as is_durable: true) and call extract_frame(video_url: <durable_rehosted_mp4>, time_s: 0). Treat each extracted frame as a candidate tile, not an approved tile: immediately call analyze_media on the frame URL with an OCR / visible text / no-text prompt before placement. This sourcing-time check must happen before the frame becomes a mood-board tile, PDF content-category tile, or voice-page curated tile. If frame 0 contains text, a title card, or a burned-in caption, drop the reel or try a later keyframe (for example time_s: 1.5 or time_s: 3) and run the same analyze_media no-text check before using that later keyframe. If no durable rehosted .mp4 is present, retry the scrape with rehost: true; do not fall back to a signed cover URL. Final PDF QA is only a backstop after this sourcing gate, never the first text-detection step.
2. NEVER crop from an IG grid screenshot. Tile widths in grid screenshots aren't pixel-aligned to clean fractions; auto-crop bleeds; result is bad. Confirmed-fail approach. If a user offers a grid screenshot, ask them for individual photos instead.
3. A mood board with zero curated tiles has failed. If neither user photos nor clean public-feed curated frames are available, unblock that first — never substitute generated stand-ins. If the user intended to provide photos and they didn't reach disk, ask them to re-upload. For start-from-scratch / no-social users, pause here and request 16+ individual photos before building visual artifacts: 6 mood-board curated tiles plus 10 separate PDF voice-page curated tiles.
4. Read the Relationship to existing grid field from Step 2. If the locked aesthetic is identical to the user's existing grid, place curated tiles as-is. If Step 2 names a slight evolution (e.g. "their grid is warm + washed, the influencer aesthetic pushes deeper shadows + cooler highlights"), apply light CSS color treatment in the HTML composite before rendering — filter: contrast(...) saturate(...) brightness(...), a subtle overlay, or a mild blend-mode treatment. Never re-pass the user's photos through gpt-image-2 — image-model re-rendering drifts face/pet/object identity and breaks the "this is actually her" anchor. If the divergence is too large for CSS treatment alone, the answer is more generated atmosphere tiles in the new aesthetic + fewer untreated curated tiles, NOT regenerating her photos.
5. Identify aesthetic coverage gaps — what part of the locked Step 2 aesthetic isn't represented yet (a specific setting, framing, lighting condition)?
6. Fill gaps with generate_image (provider="gpt-image-2") for social-content atmosphere only (load-bearing phrase — pushes gpt-image-2 away from vision-board / wallpaper energy; see Load-bearing phrases section), prompted in the Step 2 aesthetic (its palette, lighting, settings, photography style). See the hard rules below. Generated tiles go in mood-board/generated/.

🛑 HARD RULE: Generated tiles render SOCIAL CONTENT energy, not atmospheric vision-board energy.

A mood board for an influencer identity has to look like a creator's actual feed, not a vision board.

• YES: phone-camera framing, candid 3/4 angles, mirror selfies (faceless if generated), in-hand POV shots (coffee on a table, book in lap, drink at a bar), tablescape overheads, dressing-room outfit flatlays, golden-hour walking-from-behind shots, brunch food close-ups, candle-lit dinner tables, vintage shop browsing, looking-up POV with sky behind, behind-the-scenes hands-doing-the-thing.
• NO: landscape paintings, atmospheric "vibes" tiles, sunset harbors with no human signal, moss macros, pure scenery, anything that reads as desktop wallpaper or art print. If a tile looks like a Tumblr aesthetic frame from 2014 — fail. The mood board must read as a curated 2024–2026 social feed.

Prompts for generated tiles should always specify a framing (phone-camera POV, mirror selfie no face, 3/4 candid, overhead tablescape, etc.) AND a moment (the specific second of life captured), not just a "vibe."

See references/aesthetic-prompts.md "Social trend framings" for the prompt patterns.

🛑 HARD RULE: Generated tiles cover ATMOSPHERE only — never the user's real subjects.

• The user's face, body, hair → only their actual photos
• The user's dog, cat, pet, horse → only their actual photos
• Their partner, friends, family, kids → only their actual photos
• Their actual car, actual home, actual recurring objects → only their actual photos

Generated tiles render: rooms, weather, light through windows, hardwood floors, snow on cobblestones, the type of object they collect (vintage leather goods, brass watches), the kind of place they go (golden-hour skylines, woodland trails). NEVER a fake version of someone or something they actually have. Reason: face/animal fidelity drifts, and a fake version of someone's real dog reads as uncanny and insulting. The user will tell you "that's not my dog."

Composite — fill the template, don't author CSS. The board is built from a fixed template, references/templates/mood-board.template.html, rendered by html_to_png. Never ask gpt-image-2 to render a "mood board grid" — it hallucinates the tiles. You only supply data: exactly 12 .tile rows (each an image URL + a per-image background-position crop focus), the brand name, the subtitle vibe words, the font pair, and the palette tokens. The template owns the grid math, the white header band, and the dimensions once — so the cream-sliver / off-grid failures can't recur. Set {{COLS}} to 6 for portrait tiles (default; most camera-roll input) or 4 for landscape-heavy input. Use exactly 12 tiles so the title-banded board and the PDF full-bleed no-header variant fill the canvas with no empty cells.

Curated:generated ratio = ~50/50. Roughly 6 curated + 6 generated for a 12-tile board. A board that's 8+ curated tiles + 2-3 generated reads as "her camera roll dumped on a grid" — not designed. The generated tiles do the work of expanding the aesthetic into branded territory she hasn't shot yet (the brunch tablescape, the wool-coat outfit selfie, the cocktail POV). A 50/50 mix reads as a designed mood board with real-life anchoring.

Font system — locked here, propagates to Step 4 PDF.

Pick a typography pair (display + body) that matches the Visual aesthetic adjectives locked in Step 2. Don't default to Didot or Helvetica reflexively — the wrong font for the brand makes the mood board read off and undercuts the PDF. The test is whether the font's personality matches the aesthetic adjectives. Heuristic starting points:

Save the chosen pair as fontDisplay + fontBody in the persona.md Visual aesthetic section so Step 4 PDF picks them up. When fonts aren't system-installed (most of the above except Helvetica), source them from Google Fonts by filling {{FONT_IMPORT}} with a full <link rel="stylesheet" href="..."> tag.

Render the board. Fill the template tokens:

• {{TILES}} — one <div class="tile" style="background-image:url('IMAGE_URL');background-position:POS"></div> per tile. POS is the per-image crop focus: 50% 20% upper-body OOTD, 50% 35% top-down kitchen, 50% 50% centered.
• {{BRAND_NAME}}, {{SUBTITLE}} (tracked vibe words, e.g. BOSTON · WOOL COATS · COCKTAIL BARS).
• {{FONT_IMPORT}} (a Google Fonts <link> when the pair isn't system-installed), {{FONT_DISPLAY}}, {{FONT_BODY}}, {{BG}} (lightest palette tone), {{FG}}, {{MUTED}}.

Then call html_to_png with format: "png" and raster_options.viewport_px = { width: 1920, height: 1224 }. If it returns {task_id, status}, poll task_status({task_id}) until terminal and read the image URL. For the PDF full-bleed page (Step 4 page 3), render the SAME tiles through references/templates/mood-board-no-header.template.html at { width: 1920, height: 1080 } — that variant has zero gutters so the PDF page shows no cream slivers.

Mandatory checks before delivering:

1. Curated tiles required. A fully-generated mood board is a fail state. Valid curated tiles are user-supplied images or clean public-feed curated frames from a supported social handle. If you can't get either onto the board, the gate is "unblock the real curated source," not "ship without."
2. No regenerated real subjects. Re-read the hard rule above. The user's face, pet, partner, friends — never generated. Their atmosphere — fine.
3. No baked-in text on any tile. gpt-image-2 prompts must include the no-text guardrail — append "NO text anywhere in image" (load-bearing — gpt-image-2 otherwise bakes faux brand labels onto bottles, scarves, books; see Load-bearing phrases section). For handle-only reel-derived curated tiles, use frame 0 from the durable rehosted .mp4; if a reel frame still shows burned-in caption/title overlays, drop it rather than shipping text on the board.
4. Anti-stocky test. If the board reads as stock photography (uniform polish, uniform palette, studio-perfect, no real-life imperfection) — fail. Trendy 2024–2026 aesthetic energy = real-feeling snapshots > campaign-clean glossiness. The curated user tiles anchor the "real" register; generated tiles should match that energy, not slick up.
5. Palette variation within the identity. A cohesive palette is fine; a single-register board ("just brown" / "just sage" / "just beige") reads dead. Even within a warm-cognac identity, you can include warm cream, dusty rose, deep navy, sage — whatever co-exists naturally in their world.
6. Cast diversity on any generated tiles featuring people. Name an ethnicity per prompt (load-bearing procedural rule — gpt-image-2 defaults to lighter skin tones otherwise; see Load-bearing phrases section). (Note: with the hard rule above, generated tiles featuring people should be RARE — strangers in scenes only, never the user.) See references/aesthetic-prompts.md for the cast-diversity prompt pattern.
7. Zoom-read the composite. Open the PNG. Verify no overlapping tiles, no broken aspect ratios, no misaligned content. Tile borders clean.
8. Anti-template check. Should feel like THIS person, not a one-word aesthetic preset (NOT "clean girl beige", NOT "dark academia", NOT "cottagecore" unless the user named that aesthetic). If you can label the board with one TikTok aesthetic word — push for specificity.

Deliver in chat: "here's your mood board — [link to PNG]. the curated tiles are pulled from [sources]; I filled gaps in [named gaps] with generated tiles. does this feel like your visual world?"

🛑 Wait for explicit approval before moving on.

Step 4 — Influencer Persona PDF (HARD GATE — approval required)

The deliverable is a multi-page branded PDF titled [Name]'s Influencer Persona, not an inline markdown block. Never label this artifact a "media kit" — the user came here to define their persona, not to pitch partners yet; calling it a media kit is premature and undercuts the glow-up framing. The PDF lays voice-bank content into a designed document alongside the persona overview, content categories, approved mood board, and the Key Next Steps roadmap from Step 2's strategic conversation. The user approves the voice bank by approving the PDF.

What the PDF is for: the personal artifact that captures who the user is online — what they post, how they sound, what their feed looks like, and the concrete next steps to actually become that persona. It has to be well-designed — strong typography, clear hierarchy, generous white space, on-aesthetic. A sloppy PDF undercuts the entire kit.

4a — Draft the voice bank content

Generate the voice-bank text first (the same content the old skill produced inline). Save it as ./tmp/[handle]-influencer-kit/voice-bank.md for downstream skill consumption. Content sections:

• Bios (3 variants): IG main (short tagline), IG alt (longer, more specific), TikTok / casual.
• Caption modes (5): Hook, Story, Casual selfie, Vulnerable, Promo. Each mode = a CONTEXT label ("use when…") + 2 sample captions showing range (e.g., one short + one slightly longer).
• Hook openers (5): opening lines for Story / Reel / TikTok intros.
• DM + comment voice: reply-to-compliment, decline-a-brand-pitch, reply-to-hater (or "doesn't engage"), default emoji palette.
• Do & Don't (5 each): person-specific, not generic ("use specific place names" beats "post consistently").

Quality bars on the copy itself:

• Seed every caption from the scraped captions in Step 1 — sentence rhythm, vocabulary, emoji habits, recurring phrases, hook style. Chat-voice fallback only if scrape failed.
• Every caption passes the anti-generic test (could it belong to someone else? rewrite if yes).
• Every caption passes the AI-creator-voice ban list (no "apparently i [did thing] now", no em-dash parentheticals for wit, no "[x] + [y]" lists with plus connectors, no "fake errand", no listicle minimal poetry, no "I'm just a girl who…").
• Forbidden words from the identity never appear in any sample caption.
• The 2-captions-per-mode range shows genuine variation.

4b — Lay it into the PDF

PDF spec:

• Page size: 16:9 landscape, 1920×1080 per page.
• Page count: exactly 12 pages maximum, each focused on one idea. Do not expand the kit past 12 pages: the required analyze_media PDF QA path uses all_pages: true, and that sync path supports at most 12 PDF pages.
• Typography: the fontDisplay + fontBody pair locked in Step 3. NO third font, no mixing of register.
• Palette: the Visual aesthetic palette from Step 2. Page backgrounds default to the lightest tone; accents from the rest of the palette.

Page sequence (target):

1. Cover — name in display font (very large, 120–180pt), tagline subtitle in tracked body-font caps, one hero image (full-bleed or large-cropped), small "Influencer Persona · [year]" label (never "Media Kit")
2. About / Identity — 1–2 paragraphs synthesized from identity.md (real-you + influencer-you, authenticity dial named). Side imagery: 1–3 curated photos + a small stats grid (Based / Niche / Status / Voice)
3. Mood Board — FULL BLEED — the approved mood board embedded edge-to-edge, no header band, no padding. Pre-render it by filling references/templates/mood-board-no-header.template.html; the template owns the no-header grid math: 6 cols × 320w × 2 rows × 540h = 1920×1080 exactly, no gaps. Tiles touch directly with no cream between them, otherwise the PDF page shows 2–4px cream slivers on the edges and between rows. The title-banded mood board (with gutters + header) is the standalone disk artifact at mood-board.png; the edge-to-edge variant is mood-board-no-header.png and only used in the PDF.
4. Content Categories — 4 cards laid out in a 2×2 grid, each card = image + category name + one-line description + 3–4 example post types. If a scrape exists, categories are derived from the user's actual scraped feed: tally the user's 30 scraped post captions, group by topic/format, pick the four highest-volume buckets. If there is no scrape/no social yet, label the page Proposed Content Categories and derive the four cards from the user's answers, taste URLs, reference creators, and real photos; do not claim they are existing feed frequencies. Examples: Outfits & Styling / Boston Lifestyle / Home & Recipes / Brand Partnerships for a NE-coastal lifestyle creator; Workouts / Recipes / Workouts-to-Recipes / Brand Collabs for a fitness creator; Tour & Travel / Studio / Real Life / Brand Collabs for a musician. Use the user's own specificity (e.g. "Nutcracker, holiday markets, Snowport" not "city event content"). This is more useful than a generic color-swatch page. Do NOT use a flat color-swatch page or a textured-material-swatch page — both variants read as boring; the texture variant specifically reads as "all cloth" and doesn't tell anyone what the c

…(truncated)