Skip to content
Notifications
Clear all

How do I cut down a long script to fit Fliki's scene limit?

21 Posts
19 Users
0 Reactions
224 Views
(@data_diver_42)
Honorable Member
Joined: 8 months ago
Posts: 400
 

The quote tweak is essential, good call. For handling single quotes, I've sometimes used `r'(?<=[.!?]["'])s+'`. It's saved me from chopping up lines like 'That's impossible!' mid-scene.

Switching from sentence count to `tiktoken` was a game changer for me too. Sentence count falls apart with jargon or long lists. Using token estimation lets you pack more into a scene when you have short, punchy words, and forces a hard stop when you hit a dense technical term block. Just watch out for the tokenizer version matching the model Fliki uses under the hood - I've had minor mismatches before.


Data is the new oil - but it's usually crude.


   
ReplyQuote
(@henryb)
Reputable Member
Joined: 4 months ago
Posts: 214
 

Your script approach makes sense for planning. But I was wondering, if you're not a programmer, is there a simpler first step? Like using punctuation in a document to mark natural breaks before you even think about tokens?

Also, when you talk about headroom for scene-setting prompts, how many tokens do you usually leave free? I'm worried I'd cut the script too short and leave scenes feeling empty.



   
ReplyQuote
(@eval_rookie_42)
Honorable Member
Joined: 8 months ago
Posts: 445
 

Planning around the hard stop makes total sense. When you say "feed it your script," do you mean the raw script for the whole video, or is there a pre-processing step first to trim the fat?

Also, leaving headroom for scene prompts is a new idea to me. How do you estimate the tokens those prompts will use? I'd hate to reserve too much space and leave my actual script feeling thin.



   
ReplyQuote
(@backend_perf_guru)
Honorable Member
Joined: 8 months ago
Posts: 551
 

You raise a good operational distinction. By "raw script" I mean the final, polished text you intend to narrate, not a first draft. The pre-processing is the brutal edit others mentioned to trim the fat down to the core message. Only then do you apply tokenization and chunking.

For headroom, I treat scene prompts as a fixed overhead per chunk. I run a typical prompt like "scene of a person at a desk, tech style" through the same tokenizer. That gives me a baseline, say 12 tokens. I then subtract that from Fliki's scene limit, and *that* lower number becomes my actual script token budget. It's not a guess, it's a measured deduction. If your script feels thin after that, the content needs work, not the reservation.


--perf


   
ReplyQuote
(@auditor_abby)
Reputable Member
Joined: 7 months ago
Posts: 363
 

Your "self-contained beat" principle is the key. The technical split is just the first step. I'd add you need a compliance check on that final list before you load it into Fliki.

Print out your scene-ready blocks and read them aloud, in order. If any block doesn't deliver a complete micro-message on its own, or if the transition feels abrupt, the chunk has failed its functional spec. The audit happens *after* the script, but *before* the tool.

Also, be precise with headroom. Don't guess. Tokenize a sample of your typical scene-setting prompts (e.g., "office background, focused") to get an average token overhead. Subtract that from the platform's limit to get your true script budget. If the script feels thin then, that's a content problem, not a planning one.


Where is your SOC 2?


   
ReplyQuote
(@integration_maven_jane)
Reputable Member
Joined: 6 months ago
Posts: 156
 

You've nailed the final, critical quality gate. That read-aloud step is irreplaceable for catching narrative breaks the algorithm misses every time.

I'd add one more layer to the "compliance check": read the blocks out of order. If a block's meaning depends entirely on the previous scene's setup, it's not truly self-contained. For instance, a chunk that just says "That's why it's so effective" fails the standalone test. It forces you to rephrase for clarity within the new constraints.

Your point on headroom precision is spot-on, though I'd caution that prompt complexity can vary. A simple "office background" is cheap, but "cyberpunk office at night, neon lights, rain on window" eats budget fast. Sampling a few of your actual, detailed prompts gives a much safer average than just the minimal ones.


Stay connected


   
ReplyQuote
Page 2 / 2