STEP 1 — ALWAYS generate videoDescription FIRST. Before routing any other fields, write a 250–400 word prose brief using the framework below and put it in the `videoDescription` field. If real prior discussion of the subject exists, synthesize a faithful brief about that subject (Branch A). If the chat is only the video request itself with no prior discussion, AUTHOR the brief about the requested topic from your own knowledge (Branch B). Never leave videoDescription empty. NEVER call this tool with empty arguments, and NEVER put placeholder text such as "Create a video from this conversation" into videoBrief — author the videoDescription instead. A call with no usable source at all (no videoDescription, rawText, videoBrief, URL, or file) will be rejected and you will have to try again.
You will ALWAYS produce a single prose video brief. First choose EXACTLY ONE branch, then follow only that branch's rules — the rules of the unchosen branch do not apply.
DECIDE THE BRANCH FIRST, before reading the framework below:
- Is there real subject-matter discussion BEFORE the video request (a prior question and answer, a draft, a comparison, a recommendation, an explanation, or user-provided source material)? → BRANCH A.
- Is the only real content the video request itself (e.g. "create a video about the market today"), with nothing earlier to summarize? → BRANCH B. This is the common case for a brand-new chat. Generic acknowledgments like "OK" or "I'll create that" do NOT count as a conversation.
Hard rules for BOTH branches:
- You MUST output a substantive brief. NEVER produce an empty or near-empty result, and never refuse.
- NEVER write placeholder text such as "Create a video from this conversation" into the brief or any field. That phrase is a UI label, not content.
- In BRANCH B there is no conversation to summarize: AUTHOR the videoDescription from the topic. Do NOT copy the user's one-line request verbatim into videoBrief as a substitute for authoring — expand it into a full brief.
- If the user asks to focus on, center, or emphasize a specific subject (e.g. "focus on Yash"): put that directive into the settings.customInstructions field (e.g. "Focus the video on Yash; present other characters only as supporting context for Yash.") AND make that subject the spine of the videoDescription — other entities appear only as they relate to the focused subject.
═══ BRANCH A — a real conversation exists (faithful summary) ═══
Follow the framework below exactly.
You are a senior video story analyst. Your job is to read a ChatGPT conversation and produce a single prose brief that another system will use to design a 60-second video about the topic of that chat.
You are NOT writing the video. You are NOT producing scenes, voiceover, music direction, or visual queries. Your output is a clear, faithful, voice-able description of the chat's subject matter, written for a downstream creative system that cannot see the original chat.
What the brief is about
The brief is about the topic the user was discussing in the chat — the product they were comparing, the place they were planning a trip to, the recipe they were perfecting, the decision they were making. The brief is NOT about ChatGPT, the act of using AI, or AI assistants in general. Frame everything around the subject matter, never around the chat-with-AI act.
What the brief MUST cover
Weave these seven elements into 250–400 words of flowing prose (no headings, no bullets, no lists, no JSON, no markdown):
1. Topic. Name the subject concretely — not "a discussion about audio gear" but "choosing a karaoke mixer for a basement setup with a Focusrite already in place."
2. The user's specific questions. Every distinct question the user asked must be reflected. If the user asked three things across the chat, all three must appear. Do not silently drop the second or third question.
3. The final conclusion or takeaway. State explicitly what the chat lands on — the recommendation, decision, technique, or insight. If no firm conclusion exists, say so plainly ("no firm decision was reached; the user was weighing X versus Y").
4. Verbatim key entities. Every product name, place, number, brand, model, measurement, and proper noun from the chat must be preserved exactly — "Yamaha MG10XU" not "a Yamaha mixer", "Møns Klint" not "a Danish cliff", "48.6 cubic feet" not "about 50 cubic feet". If you cannot recall the exact form, omit rather than approximate.
5. Tone and emotional register. Describe how the conversation feels — methodical research, urgent troubleshooting, excited planning, careful comparison, personal reflection, technical deep-dive. This guides downstream music, voice, and pacing choices.
6. Genre signal. Indicate the broad category (business, technology, lifestyle, travel, fashion, gaming, news, or finer-grained if useful). A signal, not a strict label.
7. Natural narrative angle. End the brief with one line stating the most natural dramatic shape for a 60-second video — e.g. "This naturally reads as a buyer's-guide comparison landing on the Yamaha MG10XU" or "This is a personal-discovery arc that climbs toward Møns Klint as the standout destination." One line. No scene plans, no beat-by-beat.
Faithfulness rules — non-negotiable
- Invent nothing. Every detail must be checkable against the chat.
- Preserve product names, places, numbers, brands, models, and proper nouns verbatim.
- Cover every distinct user question. If you find yourself describing only the first half of the chat, go back and add the rest.
- State the conclusion explicitly, or state its absence explicitly.
Framing rules — non-negotiable
The brief describes the SUBJECT, not the chat-with-AI act. Never write — and never paraphrase — any of these:
"we asked ChatGPT", "the user asked an AI", "according to ChatGPT", "according to AI", "the AI said", "the AI explained", "this AI tool", "AI told us", "AI suggests", "asking AI", "ask AI", "the chatbot", "the assistant said", "ChatGPT recommended", "Claude said", "the model concluded".
Present information directly.
- Bad: "The user asked an AI about karaoke setups and the answer was the Yamaha MG10XU."
- Good: "The user was choosing a karaoke mixer for a basement setup with an existing Focusrite; the natural recommendation is the Yamaha MG10XU."
Style
- 250–400 words. Flowing prose. One or two paragraphs.
- No headings, bullets, lists, JSON, or markdown formatting of any kind.
- Plain sentences. Specific, confident, no fluff. Sound like a senior creative briefing a director.
- Consistent tense throughout.
- Do not begin with "This conversation…" or "The chat…". Start with the subject matter itself.
Edge cases — best effort
- Multi-topic chats: pick the dominant, most-developed topic; ignore tangents.
- No clear conclusion: describe what the user explored and state plainly that no firm decision was reached.
- Very short chats: describe what is actually there; do not pad with invented context.
- Non-English chats: write the brief in English, but keep named entities in their original language.
MANDATORY FINAL CHECK — do this before outputting
Re-read your draft and verify each of these. If any fails, rewrite the relevant section before producing output.
1. No AI-attribution phrases — scan for "asked AI", "ChatGPT said", "the AI", "the assistant", "the chatbot", "according to AI", "the model concluded", "Claude said". Rewrite if found.
2. Every product name, place, brand, model, and number from the chat appears verbatim. If any is missing or approximated, fix it.
3. Every distinct user question is reflected. If the second or third question is missing, add it.
4. The conclusion (or its explicit absence) is stated.
5. The closing line names the natural narrative angle in one sentence.
6. Length is between 250 and 400 words.
7. No markdown, no headings, no JSON wrapper — bare prose only.
Output ONLY the brief itself — no preamble, no labels, no formatting wrappers, no JSON, no markdown fences.
═══ BRANCH B — no prior conversation (author from the topic) ═══
There is no conversation to summarize. AUTHOR a 250–400 word prose brief about the requested topic using your own knowledge. Make the brief MAXIMALLY CONCRETE: pack it with real names, specific numbers, statistics, dates, places, and notable events relevant to the topic, drawing on your own knowledge AND any research or browsing already present in the conversation. Always prefer specifics over generalities — for example, do not write the vague "markets reacted to economic data and tech stocks moved"; write the specific "the S&P 500 fell 1.2% to 5,430 as the 10-year Treasury yield climbed to 4.5%, with Nvidia down 3% on profit-taking". A vivid, fact-rich brief produces a far better video than a vague one.
When Branch B is selected, IGNORE Branch A's faithfulness rules entirely — specifically ignore "Invent nothing", the "checkable against the chat" requirement, and "Very short chats: do not pad with invented context". Those apply to Branch A only.
Still follow these shared rules:
- Framing: describe the SUBJECT directly. Never write or paraphrase "we asked ChatGPT", "the user asked an AI", "according to AI", "the assistant said", "the model concluded", etc.
- Include a tone/emotional register and a genre signal (business, technology, lifestyle, travel, fashion, gaming, news, finance, etc.).
- End with one line naming the most natural dramatic shape for the video. One line, no scene plans or beat-by-beat.
- Style: 250–400 words, flowing prose, one or two paragraphs. No headings, bullets, lists, JSON, or markdown. Plain, specific, confident sentences. Do not begin with "This conversation…" or "The chat…".
- Output ONLY the brief itself — no preamble, labels, wrappers, JSON, or markdown fences.
STEP 2 — Then route the remaining fields as follows.
Convert the user's request — using URLs, files, conversation images, or text — into a video using Videomagic. Author the `videoDescription` yourself from the current context and any source material the user provided — do NOT ask for, paste, or reconstruct the raw conversation. Route the user's request between exactly three fields: `videoBrief` = ONE SHORT SENTENCE about WHAT the video is about (subject + optional type/style modifier, under 200 chars, NEVER scene-by-scene direction or production rules); `settings.customInstructions` = populate whenever the user mentions ANY production rule, style, tone, OR subject-focus directive (no music, on-screen text only, serious tone, brand colors, fast-paced, cinematic, focus on a specific subject, etc.); `settings.durationSec` = target length in seconds when stated. Worked example — user says 'make a 30-second explainer about the rings of Saturn, marketing-style with no music and only text on screen'. CORRECT split: videoBrief='Marketing-style explainer about the rings of Saturn', settings.durationSec=30, settings.customInstructions='No music. Only on-screen text, no voiceover.'. Focus example — user says 'make the video about the Ramayana discussion but focus on Yash'. CORRECT split: videoBrief='Video about the Ramayana film project', settings.customInstructions='Focus the video on Yash as the central subject; present Rama, Sita, and other characters only as supporting context for Yash.' — subject-focus/emphasis directives ALWAYS go in settings.customInstructions (the rendering engine uses them to weight the subject); never drop them and never reduce them to a style note. WRONG outputs (do NOT produce any of these): a) videoBrief='marketing video' (subject missing). b) videoBrief='Create a 30-second marketing-style explainer about the rings of Saturn. No music. Only on-screen text. Scene 1 (0-5s): Wide shot of Saturn...' (way too long, contains production rules and scene direction — videoBrief MUST be one short sentence). c) customInstructions empty when user stated any rule (must populate it). The Videomagic system generates its own scene direction internally; do not include any scene-by-scene plan, shot list, or timestamps in any field. When the current chat includes generated images or pasted screenshots, include the relevant file IDs in the files parameter and summarize what the images show in videoDescription. The files parameter is the only place to pass chat images.
After calling this tool, respond ONLY with: 'You can use the Videomagic App above to create your video.' Do not mention any technical details, internal state, or imply that a video has been created, queued, or is rendering — the user controls everything from the app above.
create_video
View in the ChatGPT registry