Content Machine

content-machine

Ported from Alex Lieberman's "content machine" (10X podcast). Two hard rules from the episode:

1. The origin must be human. "The number one difference between AI content that's trash and AI content that's valuable is whether the origin is human." The AI helps with everything except the part only you can do: your take.
2. The human owns the first mile and the last mile. Pick the idea; give your real words; approve the final. Everything between can be assisted.

The way to guarantee it's not slop, regardless of how good models get: the AI never invents the point of view. It structures, sequences, and checks your words. New model tells get chased forever; a human origin doesn't.

Layers

process layer   (in git, per D:\ content repo)   the pipeline steps + skills below
personal layer  (templates/, edited over time)   voice.md  +  content-lessons.md

• voice.md: how you sound. Built one of 3 ways: (a) a voice interview the machine runs, (b) paste past writing / texts / posts, or (c) start from a borrowed voice and let feedback morph it. For this user: start from past YouTube scripts + transcripts.
• content-lessons.md: after every run the user gives feedback; it's appended here; every future draft checks it first. This is the compounding part.

Pipeline

1. Idea (human picks)

Optional sweep for candidates:

• internal: scan the last 7 days of the user's notes / vault / context/log.md for "spikes" - strong point of view, a story, emotional intensity, a lesson/framework, or unusual depth.
• external: /youtube, WebSearch across Reddit / HN / X on the topic - what's already said, what's missing. Present ~10 candidate angles. The user picks one. The rest go to a backlog file (don't lose them).

2. Interview (one question at a time)

Pick ONE interviewer persona from personas/interview-panel.md (Ferriss, Rogan, Larry King, Stern, Barbaro, Walters - each questions differently). It interviews the user about the chosen idea, ONE question at a time, pushing back on vague answers, for ~15-20 min. The user answers by voice (SuperWhisper) or typing. Capture every answer verbatim into a transcript markdown file - these are the raw material; do not paraphrase them away.

3. Research pass (optional, light)

If the take needs grounding: the white-paper / primary source, current takes in market, contrarian angles, open questions only the user can answer. Feed back to the user if it changes their view.

4. Draft in the user's voice

• Choose format (LinkedIn post / X thread / long post / newsletter / script / podcast promo). Load the matching format section from templates/formats.md.
• Draft using the user's words from the transcript. Refine flow and transitions; do not replace their phrasing with your own. Model: Sonnet.
• Check content-lessons.md before writing - don't repeat a flagged mistake.

5. Editor council (score -> revise loop)

Run the draft past the personas in personas/editor-council.md (Housel, Urban, Puri, Eisenberg, Perell, + a slop detector). Each scores /10 with specific notes.

• score < 9 -> one revision incorporating the notes -> re-score once.
• score >= 9 -> done.
• One revision loop, not endless. (Same rule as the review gate - recursive loops produce worse output and never ship.)

6. Repurpose (optional)

Turn the finished piece into N derivatives (thread from a post, etc.). Each derivative goes through step 4-5 with its own format skill. Not a copy-paste.

7. Feedback -> lessons

Ask the user what the machine did well / badly this run. Append to content-lessons.md. Polish a Fable or Opus pass only if the user wants it.

Output to the user

IDEA: <the angle chosen>   (backlog: N other angles saved)
INTERVIEW: <persona>, <N> questions, transcript at <path>
DRAFT: <format>, <word count>
EDITOR COUNCIL: <score> (was <first score>)  - <one line on what changed>
FINAL: <the piece, in a code block for easy copy>

Anti-patterns

• Generating the take instead of eliciting it. If the user didn't say it, it's not their content.
• Paraphrasing the user's words into "better" phrasing - kills the voice.
• Endless editor-council loops. One revision.
• Publishing/scheduling without the user's explicit approval of the final text.
• Losing the non-chosen ideas - always save them to a backlog.