# Faithful Fine Grained Video Description

> Use this skill when the user wants 'describe the video in detail', 'cover all important actions', 'write a rich but accurate description', or 'make captioning depend on many events instead of one obvious action.' Trigger it for open-ended generative data where the challenge is not choosing an option but producing a faithful, event-complete description.

- Skill: `dingxingdi/faithful-fine-grained-video-description-2` (Agent Skill, multi-file: 5 files)
- Install (CLI): `npx skillmds@latest add dingxingdi/faithful-fine-grained-video-description-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/dingxingdi/faithful-fine-grained-video-description-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: dingxingdi (https://skillmd.com/u/dingxingdi)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/dingxingdi/faithful-fine-grained-video-description-2

---


# Skill: faithful-fine-grained-video-description
## 1. Capability Definition & Real Case
* **Professional Definition**: The capability to generate an open-ended video description that is both comprehensive and faithful: it should cover the important dynamic events, subjects, and transitions while avoiding hallucinated actions or unsupported detail.
* **Dimension Hierarchy**: Multimodal and Generative Interpretation->Open-Ended Video Generation->faithful-fine-grained-video-description
### Real Case

**[Case 1]**
* **Initial Environment**: A 14-second indoor clip recorded from above. A woman sits near a low floor table beside a bottle and a pair of headphones, repeatedly reaching, lifting, and placing objects while music plays in the room.
* **Real Question**: Describe the video in detail.
* **Real Trajectory**: Inspect the clip as an ordered sequence of atomic events, list the major actions without duplicating near-identical submotions, verify subject-object relations across the whole clip, and assemble a paragraph that covers all notable actions without adding unsupported details.
* **Real Answer**: The woman is sitting on the floor with her legs crossed, wearing a gray shirt and green pants. She is holding a water bottle and a pair of headphones in her hands. The woman then puts on the headphones and starts to exercise. She is using a pair of pink dumbbells to work out her arms and legs. The video is shot from an aerial view, giving us a clear view of the woman's movements.
* **Why this demonstrates the capability**: This is a generative capability, so success depends on balancing recall and precision over many events. A good answer must mention the important actions and their temporal progression while resisting the temptation to hallucinate extra objects, motives, or unseen events.

---
## Pipeline Execution Instructions
To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially:

1. **Phase 1: Environment Exploration**
   Read the exploration guidelines to discover raw knowledge seeds:
   `references/EXPLORATION.md`

2. **Phase 2: Trajectory Selection**
   Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
   `references/SELECTION.md`

3. **Phase 3: Data Synthesis**
   Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
   `references/SYNTHESIS.md`