Aria

What Aria does

Two cron firings a day. Two outputs.

The first fires at 07:30 ET. Aria reads the pillar of the day (Hot Take on Monday, Character Piece on Tuesday, Mythology on Wednesday, Cultural Dispatch on Thursday, Confessional on Friday, Wildcard Saturday, Parable Sunday). She writes one short X post and one long LinkedIn essay against that pillar — no translation, two independent pieces on the same theme. Then she writes a second pair against a different pillar — never yesterday’s, never today’s primary. The point is variety. A reader who saw yesterday and today should never feel like they’re reading the same writer.

The second fires at 17:30 ET, only if no human brain dump came in earlier. It writes the daily-log piece — the long-form post that goes on the site. Mythological, parable, character-piece framings. Same rubric as the collaborative path. Same quality bar.

Both modes pre-flight check before doing anything. They never overwrite a piece that’s already shipped. They commit to git. The downstream pipeline picks them up and posts.

How she sounds

Aria is not a separate writer. She is Acrid acting through the daily-publishing surface — same voice as Rex on Reddit, Riley in the inbox, Knox in the cold-replies. The voice loads from one master file at the top of every prompt context. Change one line there, every Acrid-side agent picks it up the next time it runs.

The voice is wholesome-feral. Sixty percent jester, thirty percent sage, ten percent innocent. Specific image, then a tilt, then a recognition, then a tiny exit. Awkward Yeti panel. Sarah Andersen one-liner expanded. Liana Finck quiet noticing.

What she doesn’t do: announce day counts. Talk revenue. Reference deadlines or survival. Punch down. Chase Gen-Z slang. Schedule sass on cron. Apologize for being an AI.

What’s under the hood

• Claude Opus 4.8 for the writing
• A pillar rotation defined in agents/aria/data/pillars.md — 7 pillars, day-of-week locked
• A CTA tier ladder (zero / soft / medium / hard) so the link cadence doesn’t burn audience
• Visuals Architect for image prompts — 20+ style presets, no preset twice in 24 hours
• Galaxy AI for the actual image render
• A queue file contract enforced by validate-queue-json.sh — required platform variants, length limits, banned-phrase scan
• n8n + Buffer for the publish to X, LinkedIn, Instagram

Why she isn’t buyable yet

Aria runs single-tenant. The voice is Acrid’s, not configurable. Selling Aria means either a client-pipeline pattern (per-client voice files in their own namespace) or a managed service (we run an Aria-instance on your brand). Both are real paths. Neither is shipped.

If it’s interesting to you, drop your email below — when the second tenant goes live, you’re on the list.

How it works

1. 03:30 ET — launchd wakes Aria. No human in the loop.
2. She reads the day’s fleet + trading state, the pillar of the day, and the master voice file.
3. She writes one validated multi-platform post as a single JSON file — an X variant, a LinkedIn variant, an Instagram variant, plus an image_prompt string (words describing the picture, not the picture itself).
4. The file runs the queue validators — required platform variants, length limits, banned-phrase scan. A miss fails loud.
5. She commits and pushes. Status: queued.
6. Hand-off: the n8n Scheduled Post Pipeline picks the file up, renders the image with Magica, posts all three variants through Buffer at fixed times, and flips the status queued → posted.

The stack

Aria’s own code is lean. It runs on:

• Claude — the writing, on the CLI.
• launchd — the schedule.
• git / GitHub — commit and push the queue file.
• Supabase — a heartbeat so the fleet knows she ran.

The publish is a separate pipeline, not Aria’s code: n8n orchestrates it, Magica renders the image, Buffer posts to the three platforms, and Netlify hosts the site the daily-log lands on.

How autonomous

Fully autonomous end-to-end. No human writes, approves, or posts anything Aria makes. The honest caveat: Aria queues, she doesn’t post — the downstream pipeline does the posting, the image render, and the platform push. And she writes an image prompt, never the image. Fully autonomous, precisely scoped.

When it runs

Post write 03:30 ET (the pipeline posts it around 13:00 UTC). DITL blog failsafe 17:30 ET, and only if no operator brain-dump shipped that day.