A pipeline Chief of Staff agent on the VPS decides when each of ten agents runs, is the only voice in Slack and owns the approval gate. It finds the root cause of a failure and proposes the fix. Nothing publishes without Mphoka's check mark, and no pipeline file changes without it either. Click any box to open it.
This pipeline is a standalone system beside the cockpit. It is not routed through the shared coordinator, is not a registered cockpit capability and is not in the brand registry. This page describes it as built on 11 October 2026 and contains no credentials.
Claude Code on the VPS under a dedicated, unprivileged user. It is not the cockpit's Chief of Staff coordinator, and it is separate from the YouTube pipeline agent.
Only the Chief of Staff posts and reads. No agent holds the Slack token.
Scans a watchlist of about 30 small AI accounts on Instagram, TikTok and Facebook for posts that beat their account's normal reach (5 times the median). Writes candidates with the full media link and the caption.
Pulls AI news from feeds and lab blogs and keeps the most relevant stories. The Chief of Staff starts it only when Scout came up short.
Breaks each candidate into hook, structure, pacing and why it worked. Keeps the skeleton only, never the original wording or visuals. Transcribes video with ElevenLabs.
Writes an original Cold Take on the skeleton in the brand voice, with a verdict: WORTH IT, SKIP IT, WAIT or OVERHYPED.
Draws slides from one fixed template (near-black, cyan accent, DM Sans, 1080 by 1350) using the Copywriter's words exactly. No image AI, so text cannot be misdrawn.
Renders reels with Blotato's AI video and voice template, then downloads each video and checks the format and the captions against the script.
Post name, hook, full caption, the media and a link to its Notion record. Posted by the Chief of Staff right after a render.
The Chief of Staff reads the decision, records it on the post in Notion and writes a work order. It also runs in the morning and evening checks.
Re-checks the order against Notion, then schedules the post to both accounts at the next free slot. It decides nothing and never touches Slack.
Published through the self-hosted Postiz on the same VPS, at 12:00 or 18:00 Brussels time.
Pulls reach, views, saves, shares, likes and comments at 24 hours and 7 days from Meta, with Postiz analytics as fallback. A failed read stays blank; a real zero is written as zero.
Ranks the measured posts and writes 3 to 6 directives for Scout. Skips the week when fewer than 5 posts are measured.
Rebuilds the Watchlist toward 30 active accounts. It holds 24.
Dated rules written by the Chief of Staff from evidence: operations, approval patterns, and performance once 20 posts are measured. Every agent reads it first.
Brief in #coldtake, then the approval gate and Metrics. Signal on Mondays, Scout Monthly on the 1st.
When Mphoka reacts or replies on an approval message, the approval gate is queued and the Chief of Staff handles it.
The Chief of Staff starts Scout. After each review it picks the next step: News Ingest if needed, the DNA Analyst, the Copywriter, then the Renderers only if there is something to render.
Approved posts publish at the next free slot, at most two a day.
Brief: what ran, what each step produced, whether today's posts went out. Then the approval gate and Metrics. No chain step starts after 22:30.
Write to the Chief of Staff in plain language, as a new message or in any of its threads. It answers in the thread within about two minutes and decides itself what to look up, run or investigate: Notion, logs, the queue, credits, the web, any of the ten agents. It can approve, pull or move a named post on your word, and edit a post's text before it is approved. Code and instruction changes still come back as a change proposal for your ✅.
On a post not yet approved it also counts as the approval. A post already queued for later is removed from the queue first.
Removes the queued post and schedules it at that Brussels time. An unclear time is asked back, never guessed.
Removes the post from the queue and closes it, so an old check mark cannot bring it back.
Checks in Postiz what actually exists, then schedules only the account that has nothing.
When an agent fails, gives wrong output or an alert keeps coming back, the Chief of Staff reads the logs, the agent's report and instructions, the scripts and the Notion rows, and names the first thing that went wrong with the exact line as evidence. A one-off outage is reported, not fixed. If it cannot find the cause it says what it checked and stops.
The smallest fix, one root cause per proposal. A checker refuses files outside the scope, unchanged files and syntax errors, then builds the diff and a fingerprint. The proposal is posted with both. At most three open at a time.
Within about five minutes of Mphoka's ✅ a separate applier checks that the ✅ is his, the message is unedited and the files still match the fingerprint. Then it applies the change, keeps a backup and locks the files again. A reply undo within seven days rolls it back; no answer in seven days expires it.
Can change: its own instructions, the ten agents' instructions, the run and review scripts, the approval doorbell, the DeepSeek test and the renderer code. These 19 files are locked; no agent and not the Chief of Staff can edit them directly.
Cannot change: the posting gate and Postiz, keys, timers, the dispatcher, the applier itself, other brands or the cockpit. For those it describes the problem under NEEDS YOU.
A small local service holds the Postiz key and the Meta token. No agent holds either.
The Publisher marks a post as being scheduled before it calls Postiz. A post that keeps that mark is touched by nobody until Mphoka replies retry.
The cloud copies of the agents were deleted on 11 October. If a run fails nothing covers it, so the Chief of Staff alerts in #coldtake at once and you will notice a missing 08:15 brief.
On 10 October three candidates gave one draft. Options: let News Ingest add stories every day, or lower Scout's 5 times bar. Decided together with the DeepSeek result.
The DNA Analyst picks reels for most cards, against a mix that asks for carousels and other formats too.
None has been used on a real post yet.
The Chief of Staff alerts below 3 GB. The agents and their logs live on the apps and logs volumes.
An agent's alert reaches Slack when its run and review end, which can be up to about an hour after it happened.
The agents and the Chief of Staff share one server user. The agents are not given the Slack token, but a hard separation would need a second user.