Skip to content
C.W.K.
Stream
Lesson 03 of 04 · published

The Marginal Instance Should Cost Almost Nothing

~12 min · marginal-cost, derived-work, scaling, founding-rules

Level 0Cold Workshop
0 XP0/35 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete

A Founding Rule, Restated Twice

The workshop's first project established a rule about handmade work: the Nth entry in the roster costs one line of data plus a re-render. The video is derived from data, not handcrafted frame by frame. Add a subject, add a line, rebuild.

When the series arrived, the same rule was moved one axis over: the Nth episode costs one identifier. Not a new directory, not a copied script, not a fork of the pipeline with small edits. One pointer, and the format does the rest.

Both statements are really about the same quantity — the marginal cost of the next instance — and holding that quantity near zero is what separates a tool from a pile of similar work.

Why an Episode May Not Have a Directory

There is an explicit prohibition attached: an episode never creates a project directory. It sounds like housekeeping and it is structural.

Give each episode its own directory and, within about three episodes, each one has a slightly edited copy of the format. Episode two needed a different opening, so its copy diverges. Episode three borrowed from two rather than from the original. Now there is no format — there are N variations with a shared ancestor, a fix has to be applied N times, and the pipeline has quietly degraded into a template that people copy. This happens without anyone deciding it, which is exactly why it is forbidden by rule rather than discouraged by preference.

The Real Test Is Whether Improvements Flow Backward

Here is the property that makes low marginal cost worth the discipline. When the pilot run found problems, twelve of its lessons were fixed in the tool rather than in that episode's data — a geometry check, a cache correction, a measurement that had never been taken per language. Because there is one pipeline, every one of those fixes applies to every future episode automatically.

In the copied-directory world, those same twelve fixes would have landed in one episode's copy and stayed there. The next episode would start from a different ancestor and rediscover several of them. That is the actual cost of a high marginal price: not the extra work of setting up instance N, but the fact that improvements stop compounding. A pipeline gets smarter every run; a template gets copied every run.

Code

The same rule on two axes·text
PROJECT AXIS (the founding statement)
  the Nth roster entry = one line of data + a re-render
  -> the video is DERIVED from data, never handcrafted

SERIES AXIS (the same rule, moved)
  the Nth episode = one identifier
  -> the episode is DERIVED from the format + a pointer

WHAT IS FORBIDDEN, and why it looks harmless:
  pipelines/pippalog-episode/
    episodes/ep-01/pipeline.py     <- "just a small tweak for this one"
    episodes/ep-02/pipeline.py     <- copied from ep-01, diverged
    episodes/ep-03/pipeline.py     <- copied from ep-02, diverged more
  # after three episodes there is no format, only descendants.
Where a lesson lands, and what that decides·text
the pilot run surfaced 116 rejections. each one had to be routed:

  fixed in the EPISODE's data     -> helps this episode only
  fixed in the PIPELINE           -> helps every future episode
  fixed in the TOOLKIT            -> helps every project too

12 were promoted into the pipeline or the toolkit, including:
  the subtitle-band collision check
  the invented-day-boundary check
  per-cue measured audio gain
  content-keyed plate and clip caching

# this routing is only POSSIBLE because there is one pipeline.
# with per-episode copies, every fix is local by construction.

External links

Exercise

Pick a repeated deliverable your team produces — a report, an onboarding setup, a deployment, a customer-specific configuration. Write down the marginal cost of the next one in concrete terms: files created, lines edited, decisions re-made. Then find the last improvement someone made to one instance and check whether it reached the others. If it did not, that is the compounding loss, and it is the number worth quoting when you argue for consolidation.
Hint
The clearest evidence is the same bug being fixed more than once in slightly different places. Search your history for a fix you recognize appearing in multiple files with different wording — each occurrence is somebody rediscovering a lesson that a shared implementation would have given them for free.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.