← All Projects
Live · recipe #1 in production

Content Engine

A recipe-based AI production system built on one principle: a methodology figured out once should never be re-derived. Content types are captured as recipes, brands are layered as profiles, and the marginal cost of the next piece keeps falling.

The Problem

Every content workflow I'd built before, including a full publishing engine (see its predecessor), shared one failure: the method lived in heads and chats. Each new video meant re-explaining the rubric, the voice, the structure. The insight worth building around was that the expensive thing isn't writing a script, it's the methodology behind a kind of script, and that should be a reusable asset, not a conversation you repeat.

What I Built

A master engine where each content type is a recipe: a runbook plus prompts plus the pipeline stages it emphasizes, organized in an inheritance tree so depth lives at the leaves. Brands are profiles layered on top, holding their own voice, data, and outputs while referencing the master's recipes for the how. A natural-language router resolves a request like "essay for Fizz on X" into recipe plus profile and runs a Sense, Think, Act, Learn pipeline. The Learn stage is the point: every run can crystallize new reusable beats, so the engine compounds. Judgment stays human; the mechanical layers (scoring, validation, formatting) run as deterministic code.

Recipe #1 · the video-essay method, by the numbers

01SenseOutlier scanner: 725 videos scored across an 8-channel cluster · 77 outliers flagged (score = views ÷ channel average)
02ThinkDeterministic v2.0 scorecard: 49 slogans fully scored across machine / people / discourse / speaker dimensions, regenerated from YAML by code
03Act5 script frameworks (hook-payoff · documentary · steelman-demolish · why-explainer · thesis listicle) · 65,000+ words drafted
04LearnReusable beats captured per run · an 8-pattern rhetorical toolkit shipped to the audience as a companion resource
05GovernEvery score carries a one-line evidence annotation · contested claims present both readings · the analyst's lens is disclosed

Tools & Stack

ClaudePythonYAML Data LayerDeterministic ScoringRecipe InheritanceYouTube Data APISQLiteMaster / Instance

Key Results

  • Recipe #1 live in production for a YouTube video-essay channel: research to scored analysis to finished script package, no step re-derived
  • 49-slogan scoring database on a defensible, cross-partisan rubric with 99 more candidates staged, all machine-readable, all regenerable by code
  • The "Persistence Engine" flag: a derived metric that automatically identifies failed-but-kept-alive slogans (14 flagged), turning an editorial hunch into a queryable property
  • Master/instance separation proven: the method lives in the engine, the brand's data and outputs live in the brand, so the recipe transfers to any future channel for free
  • Built by mining three predecessors (the Ghost publishing engine, a SaaS content platform, a slogan workflow) for their best IP rather than starting from zero

Challenges & Lessons

The design decision that mattered most: decompose by content type, not by brand. A brand is just data; a methodology is the asset. Second, the compounding loop has to be explicit. The Learn stage isn't a quality check, it's the reason the engine exists: work figured out once gets captured where the next run will find it. And third, keep the human where judgment lives. The engine proposes, scores, and validates mechanically, but taste decisions stay with the author. Automation is applied per recipe to the layers that deserve it, never as a button that emits content.

← PreviousGrowth Campaign