claude liam brutalist skill duration planner
The duration planner skill sizes each explainer video beat by beat using content-type floors and measured narration timing, treating total runtime as an output rather than a target to hit or pad toward.
Every video creator eventually hits the same question: how long should this video be. The duration planner skill answers it by refusing to treat duration as something you decide in advance. Instead, duration is an output. You size the script for the content, and the runtime falls out of that decision rather than the other way around.
Duration as output, not target
The skill's core thesis, stated directly in its own file, is that duration is an output, never a target. A complex mechanism might land at three or four minutes; a definitional explainer might land at thirty to sixty seconds. Both are correct if the pacing was driven by the content. A uniform target, whether that is thirty seconds or one minute, has no basis in how people actually learn from video. Hitting a fixed number either compresses content and destroys the integration a viewer needs to follow it, or pads the video with material that adds nothing but extraneous load. Either way, learning suffers.
How the pipeline actually works
The duration planner folder is deliberately small: one skill file describing the doctrine, one short reference file with the evidence and a floor-and-ceiling table, and one advisory script that reads timings and reports back. Every beat in a storyboard carries a content type set at storyboard time, and the storyboard becomes the master clock. Once Kokoro (the narration engine) measures the actual spoken timing for a beat, the skill compares that measured narration against the floor table for that beat's content type. If the narration comes in under its floor, the skill recommends a hold. If it goes over the ceiling, it recommends a split. Total runtime is simply what falls out of applying this beat by beat. The skill reports the number and stops; it does not adjust content to hit a target.
Content-type floors, concretely
Each content type carries its own consolidation floor, the minimum time a viewer's working memory needs to register a new element before the beat cuts away. A title beat needs three to five seconds. A structural or geometric beat needs six to eight. A mechanism step needs six to ten. An equation step needs seven to twelve. If narration for a beat lands under its floor, the fix is not to shorten the next beat or speed up the voice. The fix is to add a hold.
Holds are automatic, and stay in sync
Holds are applied automatically: the scene base holds the final frame of a beat up to its content type's floor, and the compile step pads that beat's audio with matching silence so audio and video stay synchronized. This can be turned off per video with a hold-floors setting in the metadata. Both the render step and the reassembly step read from the same pacing table, which matters because rendering with one setting while assembling with another would desynchronize the final video.
Where the discipline actually bites
Once an idea has landed on screen, adding decorative motion or filler graphics just to stretch the video toward a round number like one minute is treated as a coherence violation, the same kind of extraneous load that degrades learning as compressing content too far. Padding to hit thirty seconds is exactly as wrong as compressing to hit it. The skill also explicitly rejects the commonly cited "six-minute engagement" rule, noting it comes from a watch-time finding on a MOOC platform that has not been shown to replicate in real courses, and isn't itself a learning result. The trade the skill accepts is shorter watch time in exchange for a cleaner learning schema.
Key takeaways
- Duration is treated as an output of content pacing, never as a target the script is built to hit.
- Each beat's content type has a floor and ceiling; measured narration against that table decides whether to hold or split.
- Under-floor beats get an automatic hold with matching audio silence, not a rushed voice or a shortened neighbor beat.
- Padding to reach a target length is treated as seriously wrong as compressing content to fit one.
- The often-cited six-minute engagement rule is a watch-time statistic from a MOOC platform, not a validated learning result, and the skill rejects it as a basis for pacing.
Who this is for
Anyone producing explainer or educational video content who wants a defensible, repeatable method for deciding runtime, rather than guessing at a target length and cutting or padding to reach it.
Full transcript(auto-generated, with timestamps)
[0:00]Sadi, this is Liam in for bear. Today we tear down a skill that answers a question every fellow keeps hitting. How long should this video be? The honest answer is you don't decide, the content decides. The skill is called duration planner. And here is how it actually works. The easy assumption is that duration is a target you hit 30 seconds, 1 minute 5. The skill argues the opposite. Duration is an output. You size the script and let the content sit for the time it needs to land. Everything else is production convenience or padding. A skill is a folder Claude reads first. The duration planner folder is small. One skill file,
[0:34]One short reference file with the evidence and the floor ceiling table, one advisory script that reads timings and reports. That's it. The doctrine is short because the discipline is what does the work and the discipline is refusing to treat duration as a target. Here is the pipeline. Every beat carries a content type set at storyboard time. Storyboard becomes the master clock. When Kakoro measures the narration, the skill reads those two together against a floor table. If the narration is below its content types floor, recommend a hold. If it is over the ceiling, recommend a split. Total runtime is what falls out. The skill reports it and
[1:07]Stops. First design decision, and it's the thesis. Quote from the skill file, duration is an output, never a target. A complex mechanism lands at 3 or 4 minutes. A definitional explainer at 30 to 60 seconds. Both are correct. A uniform target has no learning basis. It's a production convenience that either compresses the content and destroys integration or pads it adds extraneous load. Either way, learning fails. Second decision. Every beat has a content type, and each content type has a consolidation floor. The minimum time working memory needs to register the new element before the beat cuts. A title beat needs 3 to 5 seconds. A structure
[1:42]Or geometric beat 6 to 8. A mechanism step 6 to 10. An equation step 7 to 12. If your narration lands under the floor, don't shorten the next beat and don't speed up the voice. Add a hold. Third decision. Holds are now automatic. The scene base applies a hold floor after each beat's narration. It holds the final frame up to the content type floor. The compile step pads that beats audio with matching silence so audio and video stay in sync. You can turn it off per video withhold floors in the metadata. The loadbearing detail rider and reassemble together both read the same pacing table. And rendering with
[2:15]One setting while assembling with another distincts the reel. Here is where the skill bites. Once the idea has landed, adding decorative motion or filler graphics to reach one minute is a coherence violation. Extraneous load that degrades learning. Padding to hit 30 seconds is exactly as wrong as compressing to hit it. And the six-minute engagement rule everyone quotes. That's a watch time finding from open muk. It doesn't replicate in real courses and it isn't a learning result. The trade the skill accepts shorter watch time. Clayar schema. The verdict. The duration planner skill is small on disk and long on discipline. Content type drives the floor. Kakoro drives the
[2:49]Clock, holds land, the beats, splits, break up over full ones. Total runtime is a byproduct. The skill reports it and stops and it refuses to pad or compress to hit a number production wanted for its own convenience. Your turn. Paste this into clawed code inside your own real folder. Run the duration planner skill on my beat sheet and timings. For every beat, tell me the content type, the floor, the measured narration, whether it needs a hold, and whether any beat is over its ceiling and should split. Report total runtime as an output and do not edit the sheet. Then check three things. Are the below floor beats
[3:19]Getting holds, not sure next beats? Is any overseeing beat one idea or two? And is anything padded to hit a target? If yes, cut the padding. Don't ship the coherence violation. That was the duration planner skill. Lay them in for bear.
