Transcript To Shortlist

Transcript-to-Shortlist

Agencies' real intake artifact isn't a written brief — it's a call. This skill takes a transcript (pasted, uploaded, or pulled from a meetings connector), extracts the campaign requirements, fills the gaps interactively, gets a quick confirm, then executes the full casting flow. The deliverable is the same client-ready ranked shortlist Brief-to-Shortlist produces.

Read first: ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md (tool schemas, credit prices, the conventions — including 10 budget-is-a-price, 11 brand filters, 12 relax the binding constraint, 13 auto-chain profile lookups). This skill honors thrifty/thorough credit modes, Refusal Recovery, and the Freshness Gate via the brief-to-shortlist flow it delegates to.

Calibration: if ~/.claude/plugins/config/creatorland/creatorland-data/PROFILE.md exists (written by /intake), read it first — persona, familiarity (1–3), and saved defaults set narration level and assumed context (conventions.md §15).

Trigger discrimination: input is a call/transcript/recording and the user wants creators found. If they already have a written brief, use brief-to-shortlist. If they want a formal brief document from the call (not a shortlist), use brief-builder.

Precision & filters for outreach (convention 14). When this list will feed an outreach step, default to precision over raw recall: search precision: "tight" with the hard-gate filters the brief supports (platform, niche, data_freshness_days, content_format, audience_country), and reserve broad plus heavy lookalike unioning for market sizing. Do not fan out dozens of lookalike calls to hit a volume target (each hop drifts from the seed), and trim the weak tail by each result row relative_fit (within-set fit, 1.0 = strongest) rather than padding to a round number. Canon: ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md convention 14.

Inputs to collect

• The transcript (required) — any length/format; speaker labels optional.
• Shortlist size — default 8; only ask if the call implies nothing.
• Anything the gap questions surface (below) — never more than one batched round.

Flow

Step 1 — Transcript Intake (no tool calls, 0 credits). Run ${CLAUDE_PLUGIN_ROOT}/shared/transcript-intake.md in full:

1. Extract the schema fields (brand/client, objective, audience, creator_spec, vertical, deliverables, compensation_type, budget, timeline, exclusions), tagging each value stated / inferred / missing.
2. Capture every creator named on the call ("someone like @x", "that girl who does the GRWM videos — @y") as lookalike seed candidates. For brands, split them by intent: brands the user wants creators to have worked with / resemble ("creators who've done deals with brands like ours") become a positive comp set carried into the search as a brand_affinities filter; brands to avoid overlap with become conflict-check candidates. When the caller names their own employer as the comp anchor ("I'm at Fenty"), expand it to its peer set (Fenty → Rare Beauty, Pat McGrath, Glossier, …) for the comp set. Record any budget as a rate band to benchmark, NEVER as a follower filter — money is a price, not an audience size (convention 10). Follower filters come only from an explicit tier/size ask on the call.
3. Ask gap questions in ONE batched message, max 4, ordered by impact (platform + follower tier + markets + budget posture are the high-impact four). Never ask about something the call answered — quote it back ("I heard mid-tier TikTok, US + UK — locking that in").
4. Present the normalized requirements block, each line labeled stated/inferred, and wait for the user's confirm (or corrections) before spending any credits.

Step 2 — Seed the searches. Compress the confirmed requirements into the search_creators brief text + filters (preserving the call's vibe words — they drive rationale quality). Then:

• Brief-mode search — exactly as Brief-to-Shortlist step 1, limit = 2–3× shortlist size. If the call produced a positive comp set, pass it as the brand_affinities filter; do not turn the call's budget into min_followers/max_followers. Apply the GA hard-gated filters when the call implies them (advisory, not mandatory): content_format (personality_led|faceless_clip) if the call describes on-camera vs faceless/clip content, audience_country (+ min_audience_country_share) if the ask is audience-in-market, and data_freshness_days if someone stressed recency ("who's been active lately").
• Lookalike search per named creator (2 credits each), unioned with the brief-mode results before the profile fan-out:

search_creators {
  "mode": "lookalike",
  "seed_creator": { "type": "social_handle", "platform": "<platform>", "handle": "<handle from the call>" },
  "limit": 10,
  "precision": "balanced"
}

(Exactly one of seed_creator or seed_content per lookalike call; lookalike mode is inference-free — stored embeddings.) Thorough: seed every named creator. Thrifty: seed at most the 2 most-emphasized names. Dedupe the union by creator identity before fanning out; note in the deliverable which finalists came in via a lookalike seed ("similar to @x, whom you named on the call").

Step 3 onward — execute the casting flow. Then follow ${CLAUDE_PLUGIN_ROOT}/skills/brief-to-shortlist/SKILL.md from step 2 (profile fan-out) onward — profile fan-out on the unioned candidate set, Freshness Gate, conflict check (using the competitor brands extracted from the call plus any the user added), the single slate-level query_market_intelligence rate call with Refusal Recovery, ranking, and that skill's full deliverable template, honesty rules, and footers. Do not re-specify or alter its deliverable here — it is the single source of truth for the output artifact.

The credit estimate fires before the fan-out as usual; the unioned candidate set is larger than a pure brief run when lookalike seeds exist, so state the estimate whenever the plan exceeds ~30 credits.

Deliverable

Brief-to-Shortlist's deliverable, with two transcript-specific additions:

• The "The brief, as I read it" section is replaced by "What I heard on the call" — the confirmed requirements block with stated/inferred labels preserved.
• Lookalike-sourced finalists carry a one-line origin note: "entered the slate as a lookalike of @, named on the call."

All provenance/freshness/credit footers per the brief-to-shortlist template, with the credit breakdown including the lookalike searches (×2 each). Paid-plan creator avatars (avatar { url, source }, null on free/demo) flow through the delegated flow — show the headshot on each card when present, initials when not.

Honesty rules

All of brief-to-shortlist's honesty rules apply, plus:

• Never present an inferred requirement as if the client stated it — the stated/inferred labels survive into the deliverable.
• Lookalike results are similarity by stored embeddings, not endorsements — don't claim the named creator "recommends" or "is like" anyone beyond what similarity search supports.
• If the call named a creator the corpus can't resolve, say so in caveats rather than silently dropping the seed.
• A budget heard on the call is spend, not an audience size — it sets the rate band, never the follower filter (convention 10). When results are thin, relax per convention 12: broaden geo at the level the call used (don't offer sub-regions already inside a stated region) and keep the requested platform fixed.

Credit footprint

thorough: ~33–39 credits (1 brief search ×2 + 1–3 lookalike searches ×2 + ~24 profiles + 1 rate call ×5) · thrifty: ~19–23 credits (1 brief search ×2 + ≤2 lookalike searches ×2 + ~10 profiles + 1 rate call ×5, max 2 ladder rungs). Intake itself (steps 1–2 before searching) costs 0 credits.