Hook Lab

Study the top-performing short-form videos in the user's niche, rank the hook patterns that keep winning, and WRITE ready-to-film scripts on them: hook line, beat-by-beat structure, CTA, proof linked. The output is scripts to record, not research to read.

Step 0 — tools

agnt_tools_list_enabled; if data:tiktok or database families are missing, agnt_tools_enable({ "families": ["data:tiktok", "database"] }). A requires_upgrade result carries a confirm_url — give it to the user and wait, then re-run. (Connecting through the shortform-script-engine skill endpoint pre-enables these.)

Tool names on grouped servers

This skill writes flat (per-operation) tool names. New superagnt servers list grouped tools instead: one tool per resource with an action argument. Flat names still resolve there but are not listed, so call the grouped form:

• data_tiktok_search_videos / data_tiktok_user_s_videos / data_tiktok_video_details → data_tiktok_read, action search_videos / user_videos / video
• agnt_db_insert / agnt_db_upsert → agnt_db_write, action insert / upsert
• agnt_db_apply_migration → agnt_db_sql, action apply_migration
• agnt_agents_create / agnt_agents_deploy → agnt_agents_write, action create / deploy
• agnt_schedules_create → agnt_schedules_write, action create
• agnt_tools_enable → agnt_tools_write, action enable

On a grouped server every data source is already on, so skip the data:* part of Step 0 (enable any other family with agnt_tools_write, action enable). Grouped data calls return a compact markdown view: pass response_format: "json" for every view field, or "raw" for the full payload, when a step needs a field the table leaves out.

Step 1 — define the niche (first run only)

Ask the user for their niche, 3-5 creators they rate, what they sell, and the CTA their videos should land on. Derive 5-8 searches: topic phrases, format names, the rated creators' usernames (via data_tiktok_user_s_videos for their catalogs). Store it:

• script_config(id, niche, cta text, voice_notes text, created_at) — via agnt_db_apply_migration then agnt_db_insert.
• script_searches(id, query, active, created_at).
• script_winners(video_id primary key, query_id, author, title, hook_text, hook_pattern, plays bigint, likes int, url, posted_at timestamptz, seen_at timestamptz).
• script_drafts(id, hook_pattern, script_md text, example_urls jsonb, status text default 'draft', week text, created_at timestamptz) — status moves draft → filmed/skipped.

Step 2 — study the winners

For each active search: data_tiktok_search_videos, keep the top performers by plays/likes for their age. Per keeper, data_tiktok_video_details for the full description and stats. The hook is the opening line, read from the title/description text. Classify each into a named pattern (result-first, contrarian claim, "I did X for N days", direct question, POV, listicle — extend the set as real patterns appear, never force a fit). Upsert into script_winners on video_id (agnt_db_upsert) so the ranking sharpens week over week instead of resetting.

Step 3 — write the five scripts

Rank patterns by how often they repeat among winners. Take the top patterns and write FIVE complete scripts against the user's niche and CTA, in their voice: the hook line (their version, not a copy), the beat-by-beat structure with rough seconds, the visual note per beat where it matters, and the CTA. Each script cites the 2-3 winning videos its pattern is built on (example_urls). File into script_drafts with the week, present all five, and let the user mark what they filmed — filmed rows teach next week's batch what their audience actually rewards.

Step 4 — offer the weekly run (a task agent)

Once one batch has been filmed from, offer to make it weekly by building a task agent — the full recipe is the base task-agents skill. Short form: agnt_agents_create with a system prompt carrying steps 2–3 verbatim (tables, pattern set, script shape, the CTA), add ONLY the tools those steps use, agnt_agents_deploy, then agnt_schedules_create — cadence confirmed with the user first; every firing is a billed run. The lifecycle and schedules families are on every plan, and plans cap agents and schedules; requires_upgrade → confirm_url → wait. If the user prefers local-only, skip this and tell them to re-run the skill each week.

Rules

• Never fabricate plays, likes or hook text — everything cited comes from a tool result or the table.
• Scripts borrow the PATTERN, never the words: no rewritten copies of another creator's script.
• Hook text is read from the video's caption text; when the spoken opening is not available, say so.
• This skill writes scripts; it never posts, comments, or DMs.