# A starter content-machine kit

Built by Josh Cruz / Digital Monestary. Shared with R.A.D.D.

This is the anti-slop pipeline I run to turn a real thing that happened into a post that
doesn't sound like every other AI-agency LinkedIn update. It's not a product, not a SaaS, not
a prompt pack you buy — it's the actual set of prompts and rules, generalized so you can drop
in your own voice, your own facts, and your own niche and run it yourself.

## Why a pipeline instead of "just ask the AI to write a post"

Ask a model to "write a LinkedIn post about X" and you get something generic: hedged claims,
buzzwords, a hook that could belong to any company in your space. The fix isn't a better
one-shot prompt — it's breaking the job into small steps, each with one job, each checkable
before the next step runs. That's what this kit is: seven small steps, not one big magic
prompt.

## The pattern: oracle → interview → draft → validate → council → lessons → distribute

1. **`01-oracle.md`** — score this week's raw candidate ideas (a commit, a client story you
   can generalize, a hot take) before anyone spends time drafting one. Cheap triage first.
2. **`02-interview.md`** — pull the specifics out of you, one question at a time, before a
   word gets drafted. Logs every answer as a numbered, citable line.
3. **`03-draft.md`** — draft the post, with every claim bound to a transcript line or an
   approved fact, in **your voice guide here** (bring your own — see below).
4. **`04-validate.md`** — a deterministic (non-model) pass: banned words, em dashes, every
   claim actually cited. Cheap and catches what a rushed read misses.
5. **`05-council.md`** — a small panel of judges scores the draft on hook, specificity,
   voice, structure, and compliance. One hard-gates on compliance; if it fails, it ships to a
   human instead of looping forever (cap it at two revision passes).
6. **`06-lessons.md`** — every time you rewrite an AI draft by hand, that edit becomes a
   reusable rule so the next draft doesn't repeat the mistake. This is the machine's memory.
7. **Distribute** — publish it, wherever you actually post.

## The poster variant: news → verify → draft → board → publish

`linkedin-poster.md` is a lighter pipeline for reacting to *someone else's* news or tool
instead of your own experience — verification replaces the interview, and a review board
(one human, one look, before anything ships) replaces the judged council. Use this one for
"reacting to what's happening" posts; use the main pipeline for "this actually happened to
me" posts.

## How to run this with any capable LLM

None of this is tied to a specific AI vendor or app. Each `.md` file in this kit is a prompt
plus a set of rules — paste the relevant one into any capable chat model (or wire it into a
script that calls an API) and feed it your inputs. A few notes:

- **Local models are optional.** If you want to keep this fully private, you can run the
  drafting and judging steps against a local model (for example, via
  [Ollama](https://ollama.com)) instead of a hosted API — the prompts don't change, only
  where they run. A weaker local model is genuinely fine for the deterministic-adjacent
  steps (validate, the compliance-gate judge); save your strongest model for drafting and
  for the judges that need real taste.
- **The council should not be the same model/context that drafted the post.** A model
  grading its own first draft is not a real second opinion. Use a different model, or at
  least a fresh conversation with no memory of writing the draft.
- **`04-validate.md` is a script, not a prompt**, if you want it to be reliable. String
  matching and citation-checking are exactly the kind of thing a model gets wrong under
  pressure and a script never does. The logic is short enough to write in an afternoon in
  whatever language you're comfortable in.

## This is a starter — bring your own voice

Everything in this kit is voice-agnostic on purpose. The one piece we didn't (and can't) ship
for you is **your voice guide** — how you actually sound, what words you never use, how you
open and close a post. `quality-rules.md` is the tightening discipline that sits underneath
any voice; your own voice guide is what makes the output sound like *you* instead of a
well-organized nobody. Write that page before you run step 3 — even a rough one beats none.

## Files in this kit

| File | What it does |
|---|---|
| `01-oracle.md` | Score raw ideas before drafting |
| `02-interview.md` | Pull specifics out of you, one question at a time |
| `03-draft.md` | Draft, transcript-bound, in your voice |
| `04-validate.md` | Deterministic banned-word + citation check |
| `05-council.md` | Judged review, hard compliance gate, bounded revisions |
| `06-lessons.md` | Turn human edits into reusable rules |
| `linkedin-poster.md` | Lighter pipeline for reacting to news/tools |
| `quality-rules.md` | Voice-agnostic tightening rules under any voice guide |

Everything here is also zipped as `radd-starter-kit.zip` one level up, so you can grab the
whole thing in one download.