Context as Code: Inside a Markdown-First Agent Content Machine

Key Takeaways

• Split the system into a process layer and a personal layer. The process layer lives in git and is shared by the team — it's the pipeline itself, every step from "I need an idea" to "it's published". The personal layer is files on your own machine: your codified voice, and the content lessons you've accumulated. New team members get the process for free and only build the personal half.
• Keep non-code context in the repo alongside the code. Ondrej describes running a separate repo per business function — hiring, brand deals, YouTube — and requiring everyone to work in Claude Code or Codex rather than a web chat app, so the work lands as markdown other people's agents can load. Inside a project he keeps /docs/marketing, /docs/external for deployment and hosting details, and a running user-feedback file. The term they land on for the result is a "meta harness": not a codebase any more, a context base.
• An "oracle" step mines your own systems for ideas before reaching outward. It scans the last seven days across Slack, Notion, Gmail, Linear and git looking for what they call spikes — how much point of view, story potential, emotional intensity, lesson or framework, and depth a thread has. Only then does an external sweep run across Reddit, X, YouTube and Hacker News. Roughly 15 ideas come out, and every idea you don't use is written to a Notion database ("the vault") rather than discarded.
• Codify each questioning style as its own skill. The interview panel is six named interviewer personas — Tim Ferriss, Joe Rogan, Larry King, Howard Stern, Michael Barbaro, Barbara Walters — each implemented as a separate skill so the questioning style is what varies. They ask one question at a time and push back whenever an answer is vague, over about 20 minutes of text in Claude Code. The output is a markdown file holding the transcript, key stories, core insights and quotable moments.
• Use a scored council with an explicit numeric gate. Editing mirrors the interview structure: six more persona skills (Morgan Housel, Tim Urban, Sean Puri, Greg Isenberg, David Perell, plus a dedicated slop detector) score the draft. Below 9 out of 10 it goes back through a revision loop; at or above 9 it's done. A hard threshold plus a loop is what turns "have the model review it" into something that actually terminates.
• Close the loop with a lessons file. At the end of every session the human gives feedback on what the machine did well and badly, and that is written into contentlessons.md. Before writing anything in future it re-reads that file to avoid repeating the same mistake. This is the cheapest form of persistent learning available and requires no fine-tuning.
• Keep the human at the first mile and the final mile. Their anti-slop argument is structural rather than stylistic: chasing model "tells" with a no-slop skill is a cat-and-mouse game that resets with every model release, whereas requiring that the origin be human — your ideas, your words, captured in the interview — survives model changes. The machine's job is flow, transitions and editing, not generating the substance.
• Three ways to bootstrap a voice file when there's no back catalogue: let the machine run a voice interview and turn that into the voice file; feed it anything already in your voice (past text messages, emails, Slack); or start from someone else's voice file and let feedback gradually morph it into yours.
• Design against entropy on long-running work. Dan's framing of the core problem: coding agents offer speed that wasn't previously possible, but the longer a task runs the more it risks entropy and divergence from plan — so the design question is how to minimise divergence without giving up the speed. Their answer is bigger, more explicit markdown blueprints up front, since agents hold coherence across far more information than a human standup ever could.
• Single-player versus multiplayer AI is their framing for where leverage comes from: single-player is putting a strong model in front of each person with zero behavioural change (valuable, and where most organisations should start); multiplayer is reinventing a horizontal process so leverage compounds across a whole function. They note most engagements that start as an AI problem turn out to be a data problem first.

Named Artifacts & Skills Mentioned

contentlessons.md      # feedback file, re-read before every generation
/docs/marketing        # in-repo marketing context
/docs/external         # deployment / hosting details the agent may need
/last-30-days          # Matt Van Horn's social-sweep skill, used for the external idea pass