manju-fenjing-check

A storyboard quality checker for AI short-form drama, as a DeepSeek Harness plugin.

Registers one tool, storyboard_check. Give it a storyboard segment; it returns a graded report of everything that will break before you pay to render it.

分镜体检 · 10 镜 / 15.1 秒 / 台词 41 字 / 参考图 5 张（角色基准 2 张）
必改 1 · 建议 2 · 通过 6/8 项

必改｜R1 时长账｜镜头 9
  台词 7 字，按每秒 4 字要 1.8 秒，这一镜只给了 1.4 秒
  → 加时长到 2 秒，或把台词压到 5 字以内

Status: working, verified end to end against @deepseek-ai/dsh 0.2.2-rc.1 — mounted through a --patch overlay, called by the model, report returned intact. The rule engine is deterministic: no model call, no network, no dependencies.

---

The problem

Producing AI short drama means feeding a text-to-video model (Seedance, Kling, Hailuo, Veo) a shot list plus a handful of reference images, one 15-second segment at a time. Five things go wrong over and over, and you usually find out after paying for the render:

Symptom	Root cause
The face drifts between shots	Too many character reference images per segment; or a character is told to speak while absent from the frame; or the shot description restates their build and clothing, so the model regenerates them
Timing collapses	Line length does not reconcile with shot duration; the model averages a 10-shot segment into 3-second chunks
Segments don't connect	The final shot of segment N never states a freeze-frame, so segment N+1 opens on a different world
Platform rejection	Sensitive-content terms buried in the prompts
Anachronisms	A period kitchen stocked with stainless steel and ceramic tile

This plugin turns those into eight mechanical checks you can run in a second.

The eight checks

#	Check	Rule
R1	Timing ledger	Dialogue characters ÷ reading rate ≈ seconds, reconciled per shot. Shots under 2s are flagged; declared totals must match the sum of shot durations.
R2	Reference budget	Max 2 character reference images per segment. Beyond two, identity weight is split and likeness degrades. Back-only references are useless for a face and get flagged.
R3	Speaker / frame mismatch	A lip-sync directive requires the referenced image to appear in the frame line. Otherwise the model invents a face. Also catches references to images that were never declared.
R4	Appearance leakage	Shot descriptions may state what a character does, never what they look like. Restating appearance re-rolls the face.
R5	Prompt hygiene	Prop and plate images must not depict hands or face fragments (they get welded into the image and contradict the video), and must not smuggle in modern objects. Negative-prompt tails are exempt — naming what must not appear is correct usage.
R6	Freeze-frame chain	The last shot of a segment must end with an explicit freeze-frame. Missing it means the next segment starts on a different world.
R7	Knowledge boundary	Flags wording that hints a character knows more than they could at that moment. Advisory — semantics need a human.
R8	Sensitive-term scan	A 21-term list scanned across the whole segment.

Every finding carries a shot number, the offending value, and a concrete fix.

Install

The plugin is an ordinary npm-shaped package with a dsh.bundle declaration.

# from a local clone
dsh plugin --profile web add link:/absolute/path/to/manju-fenjing-check

# or from a git source, once published
dsh plugin --profile web add "github:wujiaxiao69-arch/manju-fenjing-check#main"

Restart dsh web after installing a bundle, then confirm the mount:

dsh web --dump-config | grep manju-fenjing-check

Try it without installing

A patch overlay mounts the source file directly — good for hacking on the rules:

# scratch/dev.patch.yml
- insert:
    - id: manju-fenjing-check
      name: '/absolute/path/to/manju-fenjing-check/index.js'

dsh web --patch ./scratch/dev.patch.yml

The plugin path must be absolute; a relative path silently fails to resolve.

Tool signature

Parameter	Type	Notes
storyboard	string, required	The full segment markdown: shot blocks, reference-image prompts, video-prompt section
chars_per_second	number	Reading rate for the timing ledger. Default 4
max_character_refs	integer	Reference budget per segment. Default 2

Returns { blocking, warnings, report }, where report is the rendered text above.

Test

node run-tests.mjs              # synthetic fixtures, 10 assertions
node run-tests.mjs segment.md   # also lint a real segment

No dependencies, no framework — plain Node.

Roadmap

• Deterministic rule engine over the eight checks
• DSH plugin wrapper (defineTool, inject: ['tools'])
• R7 semantics: optional LLM pass for the knowledge boundary
• Client-side panel (input pane, graded report card) instead of plain text

Related

manju-fenjing — the production-side skill: segment splitting, timing ledger, delivery-format export, four self-check scripts.

License

MIT