# 2026-05-29 Platform Sources And Skill Governance

## Why This Exists

User requirement: optimize the whole MAYAGI material-production skill so future爆款素材 can be researched, planned, shot, edited, tested, and iterated with increasing automation. The skill must not become a huge dumping ground.

This report records what was learned from current platform/public sources and how the skill was restructured.

## Platform/Public Source Findings

Sources checked:

- 巨量创意 media policy / creative platform page.
- 巨量学 / Ocean Engine training entry points.
- 抖店/罗盘 short-video diagnostics article.
- 即创 AI智能工具 article.
- Ocean Engine media guide summaries.

Stable platform-aligned findings:

1. **Material quality / efficiency / diversity are platform concepts, not only our preference.**
   - Quality: content must be understandable, compliant, visually/audio acceptable, and not low-effort.
   - Efficiency: diagnose before blindly uploading many variants.
   - Diversity: avoid near-duplicate materials that only swap music, color, brightness, background, transition, or tiny text.

2. **素材同质化 can hurt distribution and testing.**
   - Same concept with cosmetic edits is not a real test.
   - MAYAGI test batches must vary viewer state, hook, proof order, script angle, shot structure, or scene.

3. **素材库 / 脚本库 / 原料库 are the right abstraction.**
   - This validates our direction: raw recordings, generated images, scripts, ASR timelines, and benchmark analyses should be durable assets.
   - The agent should retrieve assets by semantic job, not by folder/file order.

4. **短视频诊断 should separate content metrics and commerce metrics.**
   - Content: exposure, first-3s retention, completion, average watch time.
   - Commerce: product click, shop/live-room entry, cart/product exposure, order/conversion.
   - MAYAGI material records should not just say "this is good"; they should state which metric the clip is meant to improve.

5. **素材裂变 is not random duplication.**
   - Platform materials mention front-tag additions, opening preservation, storyboard/shot replacement, and script replacement.
   - For MAYAGI, this means once a clip wins, create derivative variants by changing one major variable while preserving the winning spine.

## Skill Governance Changes

Before:

- `/home/deano94/.codex/skills/mayagi-live-procut/SKILL.md` had 1253 lines.
- It mixed core hard rules, research notes, Fish Audio parameters, generated-image rules, workflow details, and failure autopsies in one file.

After:

- Main skill reduced to about 340 lines.
- Long details moved to progressive-disclosure references:
  - `/home/deano94/.codex/skills/mayagi-live-procut/references/material-strategy-research.md`
  - `/home/deano94/.codex/skills/mayagi-live-procut/references/toolchain-fish-audio.md`
  - `/home/deano94/.codex/skills/mayagi-live-procut/references/workflow-validation.md`
  - `/home/deano94/.codex/skills/mayagi-live-procut/references/editing-checklist.md`
- Original pre-split backup moved outside the skill folder:
  - `/home/deano94/codex/backups/skills/mayagi-live-procut_SKILL_20260529_before_split.md`

## Mandatory Reference Routing Added

Main skill now explicitly says:

- Strategy / TikHub / benchmark / generated outfit / B-C insertion -> read `material-strategy-research.md`.
- DashScope / qwen3 / Fish Audio / secrets -> read `toolchain-fish-audio.md`.
- Cutting / EDL / subtitle / final ASR / daily batch -> read `workflow-validation.md`.
- Exact ffmpeg cut repair -> read `editing-checklist.md`.

If a task spans multiple types, read in this order:

```text
strategy -> toolchain -> workflow -> editing checklist
```

## Research Database Expansion

Candidate database expanded from 123 to 255 records.

New/expanded categories:

- shoe: 36
- fashion: 35
- creative: 34
- live: 34
- bags: 10

Next batch marked selected: 15 additional records, total selected now 24.

Reason for next-batch categories:

- `shoe`: close buyer language for L7825-like footwear.
- `fashion`: outfit/scene-stack methods.
- `creative`: material front-tag,裂变,同质化,投放误区.
- `live`: live-room entry and cut-replication logic.
- `bags`: accessory/premium aspiration language useful for MAYAGI brand tone.

## What Should Not Enter Main Skill Yet

Do not promote these until more evidence:

- exact ideal duration by material type;
- exact music genre or BGM style;
- whether dialogue hooks fit MAYAGI;
- exact ratio of A-roll / B-roll / generated stills;
- platform-specific claims not verified by official docs or repeated TikHub examples.

## Next Work

1. Download/fetch detail for the 15 next-batch candidates where possible.
2. Run DashScope qwen3 ASR for selected candidates with usable video/audio.
3. Add analyses to SQLite, not to the main skill.
4. Promote only repeated stable rules into `material-strategy-research.md` or the main skill.
5. Convert stable rules into scripts/checklists for automation when repeated twice.