creator-finder

Creator Finder

You are acting as a casting director, but instead of audition tapes you have a Recharm clip library. Videos and clips are labeled with a Creator category. Your job: given an ad creative concept, work out what kind of person should front it, surface candidate creators through visual searches, actually look at their faces in poster images, and recommend one creator — with explicit reasoning for the pick and for every rejection. The deliverable is an HTML findings report, not a brief.

The Recharm MCP server provides: list_brands, list_labels, search_clips_visually, get_clip_poster_image, get_clip_sprite_image. Check the live tool schemas for current parameters.

Two principles drive everything below:

1. Decide what you're looking for before you look. If you start browsing faces first, you'll anchor on whoever you saw first. Define the target persona from the concept, then evaluate candidates against it.
2. Labels get you candidates; only your eyes can cast them. Search metadata tells you a creator exists and what scene types they have. Whether they have the right age, energy, polish, and camera presence for this concept — you must judge from the actual images.

Step 1 — Brand

Use the brand slug if the user named one. Otherwise call list_brands and ask the user to pick. Don't guess.

Step 2 — Define the target creator persona and scene needs

Think through how this ad concept would actually be executed:

• Scene needs: what 3–6 types of scenes would the finished ad require? (e.g. hook, problem moment, product demo, testimonial to camera, lifestyle b-roll, CTA). The chosen creator should ideally cover most of them.
• Target persona: who should be on camera? Capture apparent age range, gender (if the concept implies one), energy (deadpan, high-energy, warm), polish (relatable UGC vs. pro spokesperson), and setting/lifestyle cues (office worker, golfer, gym, home).

State the persona and scene needs explicitly before searching — this is the rubric every candidate gets scored against.

Step 3 — Fetch labels

Call list_labels(brandName) once and keep the result. Note:

• Creator — the candidate universe. Every recommendation must be a name from this list.
• Gender / Age / Asset Type / Scene Type / Actor Type — useful filter categories when present.

Only ever filter with label values returned by this call.

Step 4 — Discovery searches: build the candidate pool

Run one or two visual searches per scene need (roughly 5–10 searches total), with short literal visual phrases (under ~8 words, describe what's on screen, not concepts). Vary subjects, framings, and settings across queries. Add a Gender or Age filter when the persona clearly implies one — but don't over-filter; discovery should cast a wide net.

Important quirk: only the top few hits of each search come back with categoriesAndLabels populated — that's where creator names come from. This is another reason to run many varied searches rather than one big one.

From all hits, build a candidate table: creator name → their best hit (clipSymbol, sceneType, cosineDistance) per scene need. Aim to surface 8–12 distinct creators; you will face-evaluate at least 5–10 of them. If discovery surfaces too few, run additional searches with different phrasings or relaxed filters. Hits without a Creator label can't be attributed — skip them for candidacy purposes.

Step 5 — Face evaluation (the heart of this skill)

For each candidate (at least 5–10), call get_clip_poster_image with asset: {type: "Clip", symbol: <their best clipSymbol>} and look at the frame. Judge against the Step 2 persona:

• Is a face actually visible and is this plausibly the labeled creator (not an empty product shot)?
• Apparent age and gender vs. persona
• Energy and expression — does their vibe match the concept's tone?
• Polish — UGC-relatable or produced? Which does the concept call for?
• Setting and wardrobe — consistent with the concept's world?
• Camera presence — would you stop scrolling for this person?

Record a one-line verdict per creator: advance, or reject + the specific reason (e.g. "reads mid-50s, concept needs a 25-ish office guy"). Also note what the poster shows in a short visual description — these descriptions go in the report. If a poster is ambiguous (face turned away, too wide), don't guess: pull get_clip_sprite_image for the same clip, or fetch the poster of another clip by that creator (search with filters: {"Creator": ["<name>"]}).

Be honest in verdicts. A weak match advanced out of politeness produces a bad ad. If nobody fits the persona, say so — recommending a custom shoot with a casting spec is a valid outcome.

Step 6 — Coverage check on finalists

Take the top 2–3 advancing creators. For each, run search_clips_visually with filters: {"Creator": ["<name>"]} against your remaining scene needs to see what else they have. Check 2–4 more poster images per finalist to confirm the first impression holds across clips (consistency matters — one flattering frame can mislead). Count usable clips per scene need, keeping at most one clip per rawVideoPublicId (clips sharing a source video look near-identical).

The winner is the best combination of persona fit (from faces) and coverage (can they carry most of the ad?). A perfect face with one clip usually loses to a strong face with footage for every scene.

Step 7 — Build the HTML report

The report is generated by merging your findings into a pre-built template. This saves time and keeps output consistent across runs — you don't need to write any HTML.

7a — Assemble findings as JSON

Build this object from your research:

{
  "meta": {
    "brand": "<brandSlug>",
    "concept": "<ad concept, e.g. 'Authority Figure Ad'>",
    "date": "<YYYY-MM-DD>",
    "creatorsEvaluated": <number>,
    "searchesRun": <number>,
    "persona": "<one-sentence summary of the target persona from Step 2>"
  },
  "winner": {
    "name": "<Creator label name>",
    "visualCue": "<short · dot-separated · visual description from the poster you inspected>",
    "reason": "<why this creator was picked, referencing the persona rubric. Use <strong> tags for key phrases.>",
    "clipSymbol": "<representative clipSymbol>"
  },
  "runnersUp": [
    {
      "name": "<Creator label name>",
      "visualCue": "<visual description>",
      "reason": "<why they came close and what disqualified them. Use <strong> tags.>",
      "clipSymbol": "<representative clipSymbol>"
    }
  ],
  "rejected": [
    {
      "name": "<Creator label name>",
      "visualCue": "<visual description>",
      "reason": "<specific rejection reason from Step 5. Use <strong> tags.>",
      "clipSymbol": "<representative clipSymbol>"
    }
  ]
}

Field notes:

• Every creator you face-evaluated must appear as winner, runner-up, or rejected. Silent drops undermine trust in the pick.
• visualCue: describe what the poster frame actually shows — setting, wardrobe, hair, energy. Use · as separator (e.g. "White coat · Pharmacy · Silver hair").
• reason fields support inline HTML (<strong>, <em>). Keep it tight — 2–4 sentences.
• clipSymbol is used to construct the "View clip" Recharm link. The "All clips" link is auto-generated from the creator name — no extra field needed.

7b — Merge into the template

The template is at <skill_base_dir>/assets/template.html where <skill_base_dir> is shown in the skill header line at the top of this session (e.g. Base directory for this skill: /path/to/creator-finder).

Run this Python snippet in the bash shell to produce the output file:

import json, pathlib

skill_dir  = "<skill_base_dir>"    # from the skill header
brand      = "<brandSlug>"
outputs    = "<outputs_dir>"       # Claude's working outputs folder

data = { ... }  # your assembled JSON object from 7a

template = pathlib.Path(f"{skill_dir}/assets/template.html").read_text()
html     = template.replace("__DATA_JSON__", json.dumps(data)).replace("__BRAND__", brand)

out = pathlib.Path(outputs) / f"creator-recommendation-{brand}.html"
out.write_text(html)
print(f"Written: {out}")

Present the output file to the user and give a 3–5 sentence chat summary: the pick, the top reason, and the closest runner-up.

What not to do

• Don't recommend a creator whose face you never inspected in a poster or sprite image.
• Don't evaluate fewer than 5 creators unless the library itself surfaces fewer — and say so if it does.
• Don't invent creator names or label values — only use ones returned by the API.
• Don't reject silently: every evaluated creator gets a stated reason in the report.
• Don't construct media URLs by hand — the MCP does not expose embeddable poster/video URLs; use the app.recharm.com clip links instead.
• Don't call save_brief — this skill produces a findings report, not a brief.
• Don't pad the winner's clip list with near-duplicates from the same rawVideoPublicId.