A Five-Episode Season: Batch Production, Portfolio Packaging, and a Short-Drama Pipeline Interview Deep Dive
Produce a five-episode season in one run, package the whole pipeline into a portfolio piece that reads in three minutes, and turn fourteen days of engineering judgment into interview answers that hold up.
Today's Goals
- Produce a five-episode season in one run, and explain how cross-episode consistency was kept
- Package this project into portfolio material someone else can understand in three minutes
- Handle follow-up questions clearly on this pipeline's architecture trade-offs, cost control, and failure handling
Over thirteen days you built a line from nothing to something that ships. Today does two things: make it produce a season in one run, then turn it into something other people can understand and you can explain. When you finish, scroll back up and tick the three goals off.
Plain-Language Walkthrough
From one episode to a season: rely on the record, not on memory
Wrap day has a fixed ritual: reconciling the continuity log. Across dozens of shooting days, who wore which jacket in which scene and whether the prop sat left or right of the table all live in that book. Nobody maintains continuity from memory; everyone maintains it from one record they all accept.
Cross-episode consistency is the same thing. Most people's first attempt at multi-episode generation copies episode one's prompts into episode two, hand-edits a few words, copies into episode three. The first two look fine; by episode four the lead's coat has changed color, and by five the hairstyle has changed too. The reason is no mystery — every copy is another opportunity for a human edit, and after five episodes the drift is visible.
The correct approach is the world-bible record built on D2: character cards, style tokens, location table, stored exactly once. Then one hard rule:
Every shot's prompt may only be assembled from "the record plus this shot's storyboard description," never hand-written.
That rule converts "stay consistent" from something requiring discipline into something structurally impossible to violate. Appearance comes verbatim from the record's appearance field, style tokens are appended to every prompt, the voice comes from the character card's voiceId — as long as the assembly function reads only the record, episode five's lead cannot change coats.
// Prompts are assembled only from the record: appearance verbatim from cast, style tokens appended
function imagePrompt(world, shot) {
const who = shot.characters
.map((id) => world.cast.find((c) => c.id === id))
.filter(Boolean)
.map((c) => `${c.name} (${c.appearance})`)
.join(', ')
return [shot.visual, who, `${shot.shotSize} shot`, ...world.styleTokens]
.filter(Boolean)
.join(', ')
}
// Cross-episode hooks: episode N's hook must equal episode N+1's pickUp word for word
function checkHookChain(season) {
const broken = []
for (let i = 0; i < season.length - 1; i += 1) {
if (season[i + 1].pickUp !== season[i].hook) {
broken.push(`E${season[i].n} -> E${season[i + 1].n}`)
}
}
return { pass: broken.length === 0, broken }
}def image_prompt(world, shot):
"""Prompts come only from the record: appearance verbatim, style tokens appended."""
by_id = {c["id"]: c for c in world["cast"]}
who = ", ".join(
f"{by_id[cid]['name']} ({by_id[cid]['appearance']})"
for cid in shot["characters"]
if cid in by_id
)
parts = [shot["visual"], who, f"{shot['shot_size']} shot", *world["style_tokens"]]
return ", ".join(p for p in parts if p)
def check_hook_chain(season):
"""Cross-episode hooks: episode N's hook must equal episode N+1's pick_up word for word."""
broken = [
f"E{cur['n']} -> E{nxt['n']}"
for cur, nxt in zip(season, season[1:])
if nxt["pick_up"] != cur["hook"]
]
return {"pass": not broken, "broken": broken}The second thing a season adds over an episode is cross-episode hooks. An episode is self-contained; a season has to make people keep watching, so each episode ends on a question the next one picks up. Do not rely on memory for that either: pin hook and pickUp fields in each episode's plan, require episode N's hook to equal episode N+1's pickUp word for word, and validate before the run starts. A broken chain raises an error rather than being discovered after all five episodes are generated.
The lab's consistency check has five items, all looking at inputs rather than pictures: every character exists in the record, voices are unchanged across episodes, appearance descriptions come verbatim from the record, style tokens are present on every shot, and hooks chain end to end. Why only inputs? Because whether the picture is good is D11's machine review; this day checks whether the inputs drifted — and only undrifted input makes undrifted output possible. The two checks are complementary and neither substitutes for the other.
And once a season has run, what do you actually hold? For that, look at the books.
Interview questions
Walk me through the AI content pipeline you built. What was the hardest part?介绍一下你做的这条 AI 内容生产线,它最难的地方在哪?
Common in ChinaCommon overseasBasic#project-storytelling#system-design🔒 Unlock this chapter to see the answer keys.
How do you keep characters and visual style consistent across many generated episodes, and what breaks at a hundred episodes?多集连续生成时,人物与画风的跨集一致性你是怎么保证的?如果要做一百集会遇到什么新问题?
Common in ChinaCommon overseasIntermediate#consistency#prompt-assembly🔒 Unlock this chapter to see the answer keys.
If you rebuilt this pipeline from scratch, what would you change architecturally?如果让你重做一遍这条生产线,架构上你会怎么改?
Common in ChinaCommon overseasDeep dive#architecture-review#trade-offs🔒 Unlock this chapter to see the answer keys.
