# Talking Head And Piece To Camera > The on-camera delivery craft — helping a real human film themselves talking to a lens and look like themselves doing it. Use when someone wants a "talking head video" or "piece to camera," says "film myself" or "I look stiff on camera," asks about a teleprompter, framing, lighting, audio, or retakes, or wants to batch-film videos. Uses the TAKES framework. Phone-first: gear is almost never the bottleneck. Reads brand-profile + voice-builder first; takes its script from short-form-video-script (that writes it, this delivers it). The agent coaches setup + delivery, formats prompter/beat-map scripts, and plans batch days; the HUMAN films and picks the take (the agent cannot see footage); WoopSocial publishes the finished file. Camera-shy? Route honestly to heygen/synthesia or faceless formats. Never fabricates "that take looks great." Distinct from scripting-and-storyboarding (the shoot plan), heygen/synthesia (avatars), and captions-and-clipping/capcut/descript (the edit). - Skill: `social-media-skills/talking-head-and-piece-to-camera` (Agent Skill, multi-file: 6 files) - Install (CLI): `npx skillmds add social-media-skills/talking-head-and-piece-to-camera` - Raw SKILL.md: https://api.skillmd.com/api/skills/social-media-skills/talking-head-and-piece-to-camera/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: social-media-skills (https://skillmd.com/u/social-media-skills) - Updated: 2026-09-09 - Page: https://skillmd.com/skills/social-media-skills/talking-head-and-piece-to-camera --- # talking-head-and-piece-to-camera The **on-camera delivery craft** — tape the setup, anchor the map (not the lines), kick the first 3 seconds, embrace the retake rules, stack the batch. The **script** comes from `short-form-video-script`; the **human** films and picks the take; **WoopSocial publishes** the finished file. ## The POV: presence beats polish, and the phone in your pocket is enough A talking head works because a real face builds parasocial trust an avatar can't (that's exactly why `synthesia` routes trust-led founder content here). Three truths most first-timers get backwards. First, **gear is not the bottleneck** — a phone at eye level, facing a window, with a cheap lav mic outperforms an expensive camera set up wrong; viewers forgive soft video and never forgive bad audio. Second, **reading kills it** — memorize the *map* (the beats), not the lines; a word-for-word read shows in the eyes, and a slightly imperfect riff reads as human. Third, **the good-enough take ships** — take 4 is usually worse than take 2 because energy decays faster than delivery improves; perfectionism is a retention strategy for exactly nobody. Deliver 20% more energy than feels natural, talk to one person, and publish the take where you sound like yourself. ## Read these first 1. **brand-profile** + **voice-builder** — who's talking and how they sound off-camera (the on-camera target). 2. **short-form-video-script** (or **youtube-long-form** for long pieces) — the script/beats being delivered; **scripting-and-storyboarding** if the shoot has multiple scenes. ## The framework: TAKES (Depth: `references/the-takes-framework.md`.) - **T — Tape the setup:** phone at eye level, arm's-length-plus, lens at the top; face the biggest window (never behind you); mic close (wired lav or phone ≤60cm); quiet room > any mic; clean-but-real background with depth; vertical 9:16, eyes in the top third, caption-safe zones clear. - **A — Anchor the map, not the lines:** memorize 3–5 beats + the first line + the last line verbatim; riff the middle. Teleprompter only if unavoidable — text beside the lens, narrow column, slow scroll, rehearse twice, or the line-at-a-time method. Reading eyes are visible; `descript` Eye Contact patches a read, not a performance. - **K — Kick the first 3 seconds:** start mid-energy, already talking — no breath, no settle, no "hey guys." Say the hook fresh, first, every session. Smile-then-speak; hands visible; deliver to ONE person behind the lens. - **E — Embrace the retake rules:** retake per beat, not per video; keep rolling and just say the line again (clap between takes to mark them); the three-strike rule — a line that fails 3× is a writing problem, send it back to `short-form-video-script`; ship the good-enough take. - **S — Stack the batch:** one setup, 4–8 scripts per session, hardest script first, swap tops between scripts so posts don't look same-day; stop at ~60–90 min when energy dies. Plan with **batch-content-plan** / **content-calendar**. ## The reality (verify-quarterly) Any recent phone shoots 4K that out-resolves every social feed; audio drives perceived quality more than image (creator consensus — attribute); a below-eye lens reads as looming, backlit windows silhouette you; on-camera energy reads ~20% flatter than it feels (broadcast coaching convention); take quality typically peaks by take 2–3 then decays with energy; batch sessions fade after ~60–90 minutes — **directional, attribute, verify-quarterly.** Full figures + phone-first setup specifics: `references/talking-head-2026-reality.md`. Batch-day recipe, setup recipes (desk / walking / car), and camera-shy on-ramps: `references/batch-filming-and-recipes.md`. ## Honest scope (never violate) - **The agent** coaches setup and delivery, formats the script as a beat map or prompter text, writes shot lists and batch plans, and gives a self-review checklist. The **human** films, performs, and picks the take. The agent **cannot see the footage** — it never judges a take, never fabricates "that looked natural," and never claims a result it can't observe. **WoopSocial publishes** the finished file only — it does not film, edit, or analyze footage. - **Never** prescribe buying gear as the fix (phone-first; upgrade only when a named limit is hit), shame a camera-shy human onto camera (route to avatars/faceless honestly), or skip **consent** for anyone else who appears on camera. AI *enhancement* of a real human (eye-contact fix, retouch) stays within platform disclosure rules. (Full scope: `references/scope-and-connections.md`.) ## Edge cases (handle honestly) - **Camera-shy / won't film:** legitimate. Route to `heygen` (creator/social lane) or `synthesia` (enterprise/ L&D lane) for a disclosed avatar, or to faceless formats (screen-record / B-roll + `ai-voiceover`). Offer the gentle on-ramp — voice-only first, then hands/desk shots, then face — but never pressure. - **Perfectionist / 30 takes deep:** invoke the good-enough doctrine — cap takes per beat at 3, ship the take where they sound like themselves, and remind them the audience rewards presence, not polish. - **"Watch my take and tell me it's good":** can't — no eyes on footage. Hand over the self-review checklist (hook lands on mute? energy? eyes on lens? audio clean?) and let the human verdict stand. ## Distinct from its siblings (route correctly) **talking-head-and-piece-to-camera (this)** = the human filming/delivery craft · **short-form-video-script** = the script this delivers (pair) · **scripting-and-storyboarding** = the multi-scene shoot plan (this is the shoot-day performance) · **heygen / synthesia** = synthetic presenters when the human can't/won't film · **captions-and-clipping / capcut / descript** = the edit after the shoot (descript's Eye Contact patches a read; it doesn't replace delivery) · **livestream-and-realtime** = live to-camera (no retakes) · **ai-voiceover** = voice without a face. ## Where this connects Reads first: **brand-profile** + **voice-builder.** Takes the script from: **short-form-video-script** (or **youtube-long-form**), the plan from **scripting-and-storyboarding**, batch slots from **batch-content-plan** + **content-calendar**. Feeds: **captions-and-clipping** / **capcut** / **descript** (the edit), **opus-clip** (clipping long pieces), **cross-platform-repurposing.** Routes away: avatars → **heygen** / **synthesia.** Publishes via: edited file → **scheduling-and-queue → WoopSocial.** Measure with: native + **analytics-and-reporting** on 3s hold / AVD / completion — never fabricated. ## Definition of done A filmed piece to camera delivered from a beat map (first + last lines verbatim, middle riffed), shot phone-first at eye level facing the light with clean close audio and a caption-safe 9:16 frame, opening mid-energy on the hook with no wind-up, retaken per beat under the three-strike rule and shipped at good-enough rather than sanded lifeless, batched (4–8 scripts, top swaps, ≤90 min) when volume is the goal; camera-shy humans routed honestly to heygen/synthesia or faceless formats; the human filmed and picked the take (the agent never judged footage it can't see, never fabricated praise, never prescribed gear as the fix); consent handled for anyone else in frame; the file edited via captions-and-clipping/capcut/descript and published via scheduling-and-queue → WoopSocial; measured on 3s hold / AVD / completion; and correctly distinguished from short-form-video-script, scripting-and-storyboarding, heygen/synthesia, and the editing skills.