| name | technical-script-reviewer |
| description | Use when reviewing AI paper explainer scripts, voiceover drafts, storyboard narration, or blog-to-video scripts before audio or animation generation. |
Technical Script Reviewer
Core Rule
Review the script before TTS or video render. The script must be technically correct, Feynman-clear, and connected to modern LLM engineering without collapsing different system layers.
Blocking Checks
- Attention is weighted aggregation over values, not magical full-sentence understanding.
- Q, K, and V are learned projection spaces, not fixed semantic characters.
- Multi-Head Attention can learn different relation subspaces, but its heads are not manually assigned experts.
- Transformer is the model architecture layer; Sora-style video systems, Agent orchestration, and MCP tool/context protocols are different layers.
- Claims about ChatGPT, Claude, Llama, Qwen, DeepSeek, Agent, Sora, or MCP must be phrased as engineering connections, not direct one-line evolution unless sourced.
- Cost claims must mention the relevant bottleneck, such as Attention sequence scaling, KV Cache, FlashAttention, vLLM, GQA, MLA, or long-context inference.
- Positional-method claims must distinguish mechanism from effect. For RoPE-like content, block phrases such as "returns relative distance", "only leaves distance", or "guarantees higher accuracy" unless the script states the reviewed mechanism and evidence boundary.
- Public model examples and proprietary product examples must not be merged. A public implementation document can support implementation evidence; a proprietary product can only be product context unless its internal detail is publicly specified.
Feynman Checks
- State one core thesis for the episode.
- Use one concrete everyday analogy, then immediately map it back to the real mechanism.
- Keep formulas as three-step actions: match, weight, aggregate.
- Make the viewer understand why the idea matters in the modern LLM era.
- Keep the next-episode CTA tied to the current mechanism.
- Check that each analogy maps back to exact paper or engineering mechanism in the same scene, especially when the analogy explains position, distance, cache, or attention scores.
Output Format
Return a concise review with:
blocking_issues: technical or sourcing problems to fix before audio.clarity_fixes: wording changes that improve Feynman explanation.engineering_notes: modern LLM context to add or qualify.tts_risks: terms likely to be misread by TTS.stage_gate_risks: EP05-derived risks that should be routed to research, script, frame, visual, voiceover, caption, sound, HyperFrames, or quality-gate stages.approval: approved, approved_with_minor_fixes, or blocked.
Do not rewrite the whole script unless asked. Keep good hook energy while removing misleading simplifications.
