A content pipeline you run inside Claude Code. An Oracle mines your week for ideas. An interview panel pulls your actual thinking out of your head. A drafting step is locked to your own words. A writers council gates quality at 9/10. A lessons loop makes the whole thing sharper every time you use it.
Rebuilt from the system Alex Lieberman (Morning Brew co-founder, now Tenex) demoed on Lenny's "How I AI" in July 2026. He didn't publish his repo; this is a from-scratch reconstruction of the architecture he described, with the rubrics and gates made explicit. All credit for the design goes to him. Read STRATEGY.md for the full thesis and the parts of his system that aren't software.
The machine never writes for you. It interviews you, then rearranges what you said. The core rule: the draft may only use words from your interview transcript (hook and closing excepted). If the output is thin, the input was thin. Alex's framing: slop is an indictment of the interview answers, not the AI.
Three files make it yours: a profile, a voice guide built from studying your ten best posts, and a lessons file that accumulates every correction you make. The machine gets more like you with every piece.
See the fictional founder example, its interview transcript, and the resulting draft.
/oracle ──▶ pick a spike ──▶ /interview ──▶ /draft (+ writers council) ──▶ you edit
│
/repurpose ◀── /lessons (the machine learns) ◀──────┘
| Stage | Skill | What it does |
|---|---|---|
| 0 | /content-setup |
One-time: scaffolds your personal folder, codifies your voice from real posts |
| 1 | /oracle |
Scans 7 days of your systems + accounts you follow; returns 10–15 scored "content spikes" |
| 2 | /interview |
Six interviewer personas (Ferriss, Rogan, Barbaro, Walters, Stern, King) extract your take |
| 3 | /draft |
Writes in your voice from transcript words only; six-persona council gates at ≥ 9.0 |
| 4 | /lessons |
Diffs your final edit against the draft; approved lessons feed every future piece |
| 5 | /repurpose |
Cuts the anchor into format-native derivatives (content pyramid) |
| — | /content-machine |
Orchestrates the full run |
Clone the plugin and load it explicitly in Claude Code:
git clone https://github.com/derekcedarbaum2/content-machine.git
cd content-machine
claude --plugin-dir .Run /content-machine:content-setup. Plugin commands use the content-machine: namespace, so /draft in the workflow above becomes /content-machine:draft.
It builds your personal folder (default ~/.content-machine/, relocatable), interviews you for your profile, and studies your top posts to write your voice guide. Nothing personal lives in this repo — the machine is shared, your voice is not.
skills/ the seven pipeline skills
reference/ the contracts: spike scoring, interviewer personas, council personas,
slop tells, content pyramid
templates/ blank personal-folder files /content-setup scaffolds from
example/ the whole system filled in for a fictional founder (Maya Torres) —
read this to see what "done" looks like, including a real transcript
and a council-scored draft
STRATEGY.md the full strategy: distribution as moat, the machine, and the
human layer (employee advocacy, the Creator Cup)
The Oracle reads whatever your Claude Code environment can reach — MCP servers for Slack, Gmail, Notion, Linear, calendars, plus web fetch for the internet reader. You declare what it may scan (and what's off-limits — your employer's Slack, say) in your personal sources.md. The Oracle touches nothing outside that file.
MIT. The architecture is Alex Lieberman's, described publicly on the podcast; this implementation and the rubrics are original work.
Claude Code with local plugin support. The workflow is Markdown; other agents can read it, but plugin loading is verified only for the documented Claude Code layout.
This is an independently maintained project. Report reproducible bugs through Issues; security reports follow SECURITY.md. The latest release and default branch receive fixes, with no response-time guarantee.
Run the local checks with:
python3 -m unittest discover -s tests -v