feat(skills): sync additional miaobao workflows

This commit is contained in:
2026-06-29 14:19:05 +08:00
parent 160a4f2fb4
commit 716d41b36e
5 changed files with 358 additions and 0 deletions
@@ -0,0 +1,82 @@
---
name: miaobao-jianying-alignment
description: Use when a Miaobao video-editing issue, spec, or implementation says 对齐剪映, CapCut, 剪映初剪, jianying-alignment, capcut-alignment, miaobao-video-benchmark, or needs real Jianying Pro behavior as the validation standard.
---
# 秒爆 Jianying Alignment
## Goal
Derive 秒爆 video-editing rules from real Jianying Pro/CapCut behavior before
drafting issues, specs, tests, or implementation plans.
Do not treat "对齐剪映" as a slogan. Turn it into observed rules, evidence,
invalid cases, and acceptance checks.
## Use When
- User says 对齐剪映, 剪映初剪, CapCut, Jianying, or asks whether 秒爆 video
editing matches mainstream editor behavior.
- A video-editing issue depends on timeline, track, clip, trim, transition,
snapping, overlap, linking, resource-library, property-panel, or export
behavior.
- Existing 秒爆 behavior feels plausible but needs a competitor truth surface
before filing an issue or writing code.
## Required Gates
- Real Jianying gate: operate Jianying Pro with Computer Use for the relevant
flow unless the user explicitly says it is unavailable. Screenshots supplied
by the user are useful evidence, but are not enough for a benchmark claim.
- Research gate: when public documentation helps, prefer official CapCut/Jianying
docs or stable public tutorials. Public docs supplement the desktop probe;
they do not replace it.
- Evidence gate: capture the actual Jianying state used for the conclusion.
Activate the app before screenshotting; discard screenshots that capture
Codex or another foreground app.
- Classification gate: mark each conclusion as 【实测】, 【资料】, 【推断】, or
【未验证+原因】.
## Probe Modes
- Semantic UX probe: use Computer Use screenshots and state inspection for
questions like "where is transition added", "what panel opens on selection",
or "does the timeline show default placeholders".
- Quantified interaction probe: when rules involve drag thresholds, snapping,
overlap handling, dead zones, or layout coordinates, use the
`probing-macos-gui` method: AXUIElement readouts first, Quartz CGEvent drags,
operation -> readout -> judgement, and bidirectional threshold checks.
- Existing local tools, when present: `~/code/real_product/miaobao/scripts/jianying-probe/`
(`ax_jy.py`, `drag.py`) and `jianying-timeline-report.md`.
## Output Shape
Return a compact benchmark block before any issue/spec:
```text
剪映验证结论:
- 【实测】...
- 【资料】...
- 【推断】...
- 【未验证+原因】...
秒爆差距:
- ...
可落地规则:
- 数据模型:
- UI 入口:
- 无效场景:
- 验收标准:
证据:
- local screenshot/report paths or uploaded asset links
```
## Handoff
- For bug capture, pass the benchmark block into `miaobao-capture-bug`.
- For implementation/acceptance, pass the rules into `miaobao-issue-acceptance`
before tests or code changes.
- Do not create/update issues or edit code from this skill without the relevant
downstream skill's confirmation gate.
+94
View File
@@ -0,0 +1,94 @@
---
name: update-miaobao
description: Review completed Miaobao work and propose updates to Miaobao-related skills. Use when the user asks to reflect on a Miaobao task, update Miaobao workflows, improve miaobao-issue-acceptance, update_miaobao, up-miaobao, or maintain Miaobao skills after an acceptance, PR, or bug-capture run.
---
# Update Miaobao Skills
## Goal
Turn feedback from completed Miaobao work into high-signal updates for Miaobao
skills. This skill only manages skills. Do not update memory, docs, issue
comments, or scripts unless the user separately asks for that.
## Targets
- Default flagship:
`/Users/zfc/.codex/skills/miaobao-issue-acceptance/SKILL.md`
- Bug capture:
`/Users/zfc/.agents/skills/miaobao-capture-bug/SKILL.md`
- Future Miaobao skills under `/Users/zfc/.codex/skills/` or
`/Users/zfc/.agents/skills/`, only after the split gate below passes.
## Quick Start
1. Identify the completed Miaobao task, issue, PR, or bug-capture run.
2. Extract pits, wins, and target skills from the conversation/evidence.
3. Prefer updating an existing skill, especially `miaobao-issue-acceptance`.
4. Present target path, reason, and unified diff before writing.
5. Wait for explicit user confirmation before changing any skill file.
6. After writing, reread the changed file and report the exact path.
## Routing
Use `miaobao-issue-acceptance` for:
- worktree setup, runtime files, ports, Hocuspocus, startup, and cleanup
- Superpowers full workflow, TDD, root-cause debugging, verification, review,
PR preparation, merge, and issue closure
- dedicated browser/profile handoff, login state, Chrome plugin/control ports,
screenshots, console/network evidence, and manual acceptance
- AI media capability verification: image/video/audio/BGM model setup, real
task creation, generated assets, and media stream proof
- Gitea PR hygiene, especially `Closes #<issue>` and post-merge closure checks
Use `miaobao-capture-bug` for:
- turning vague user notes, screenshots, or recordings into Gitea issues
- evidence extraction, media upload, version lookup, labels, priority, and final
confirmation gates for issue create/update/comment/upload
- reproducing and publishing demo videos for bug reports
If feedback spans both, propose separate diffs per skill.
## Split Gate
Creating a new Miaobao skill, or proposing a skill-composition/DAG workflow, is
allowed only when all are true:
- The same friction or workflow pattern has repeated.
- The boundary is clear enough to say what the skill owns and does not own.
- The trigger is specific enough for a future agent to choose it reliably.
- It does not substantially overlap with an existing Miaobao skill.
- The workflow has concrete acceptance evidence or deterministic checks.
- It can cooperate with the flagship skill through an explicit sequence.
If any condition is weak, update the flagship skill instead and mention the
possible future split only as a note.
## Diff Gate
Never silently edit a Miaobao skill from reflection alone. Before writing, show:
- target skill path
- why this target is correct
- exact unified diff
- whether the change fixes a pit, preserves a win, or adds a split/DAG gate
- why the lesson is high confidence
Only apply the patch after explicit confirmation. A reply of `1` confirms only
the reviewed diff/action set, not additional unreviewed edits.
## Output
Keep output compact: summarize pits, wins, target skill, proposed diff, and any
new skill/DAG recommendation with confidence and trigger conditions.
## Common Mistakes
- Editing memory, docs, issues, or scripts from this skill.
- Creating a new skill from a one-off idea instead of repeated workflow pressure.
- Forgetting to preserve wins, such as PR closure or strong browser handoff.
- Asking generic update questions when the user already gave concrete feedback.
- Mixing bug-capture rules into acceptance workflow rules.
- Applying a patch before the user has reviewed the exact diff.