arizuko › products › Content pipeline (Inari) › setup

Content pipeline setup

1. Prerequisites

A running instance (getting started).
CODEX_API_KEY — declared required. The oracle skill does the multi-source synthesis; without it the agent drafts from one source at a time.
An approval channel (TELEGRAM_BOT_TOKEN) and at least one publish adapter, if you want it to publish rather than hand you text.

2. Seed a group from it

arizuko group <instance> add telegram:user/12345678 content --product creator

3. Write the style file first

facts/style.md is the whole product. It ships as an instruction to the operator rather than as content, because a generic voice file produces generic drafts. Record: formal or casual, first or third person, what is in scope, what is off-limits, length and formatting preferences, and the per-platform rules (Bluesky brevity, thread shape on X).

Draft quality tracks this one file. If the drafts read like nobody in particular wrote them, the fix is here — not in the prompt, and not in the model.

4. Wire publishing

BLUESKY_IDENTIFIER=you.bsky.social
BLUESKY_APP_PASSWORD=...

The publish step is an ordinary adapter call, so the agent can only post where you have wired an adapter and granted the action. No adapter means no publishing — the drafts still come back in chat.

5. Verify

arizuko send <instance> content "pitch me three angles on last week's outage" --wait

A correct answer pitches and then asks for a brief and a platform. If it hands you a finished draft unasked, SOUL.md did not land.

6. Tuning

Voice — facts/style.md, not SOUL.md. The persona governs the working habits; the style file governs the output.
Approval — the agent waits by default. Removing that is a change to SOUL.md, and it is the one change worth thinking twice about.
Compose with others — blend socials into the same group for distribution; the two personas collide, so keep one persona and take the other product’s facts and rules.