AI literacy
A harness is not the model — it is the wiring: what files it reads, what rules apply, what happens next, and who owns the output.
What a harness contains
| Layer | Example in the Club stack |
|---|---|
| Context | Markdown in Obsidian · GitHub-synced second brain · files attached to a Gem |
| Rules | Mega-prompt · agent instructions · design token dictionary |
| Trigger | You paste a link · phone Flashcard · Zoom webhook fires n8n |
| Hand-off | Distill → vault note · Publish · member Dashboard · logged decision (AI Orchestrator) |
Codex itself is not “local AI” — it is interconnected Markdown members read. AI reads in when you attach a harness (2026-08-19 > Codex harness · deterministic vs probabilistic AI).
Harness vs chat window
| Chat window | Harness |
|---|---|
| Thread amnesia next week | Context lives in files you curate |
| Same question → different answers (probabilistic) | Constrained → deterministic where it matters |
| You are the entire pipeline | One person orchestrates many steps |
Examples named in the room
- Post-session pipeline — Zoom ready → webhook → n8n → thirteen named steps → session on Codex (was ~two hours manual; minutes unattended) — 2026-08-19.
- Grok Bot — “fetch avatar profile, business brand voice, influencer list from the second brain.”
- Gemini Gem + vault files — associate token or doctrine files so graphic passes stay on signature.
- Operator slash workflows — e.g. distill a vault link through a fixed Cursor command (Cursor · private infrastructure).
Business brand belongs in the harness
Brand voice is not a chat personality — it is what the company seat holds in Tripartite reins: attributable language, avatar fit, lists you trust. Wire those files into the harness (Mega-prompt · Grok Bot) before you automate outbound (Feedback as learning currency on readability).
Your move this week
Sketch three boxes — Capture · Synthesis · Execution — from AI Orchestrator. Circle the box still “pure chat.” Name one file or automation that would harness it.