Killfeed Weekly: a Call of Duty news channel run by an AI agent

What it does

• Reads the scene daily. At 7 AM PT a sweep checks the 18 sources and files every new story in an inbox table with a status (confirmed or rumored), the people involved, the source, a Short score and a "handle with care" note where a story is sensitive.
• Mines quotes from two-hour podcasts. New videos from the top channels go through a transcript step that finds the quotable moment: who said it, the exact seconds, what it's about. That's how a watch party becomes a 15-second credited clip.
• Writes a story, not a briefing. On Monday the editor reads the week's inbox, researches the people behind each story, and writes a 10-minute script where every segment has a protagonist and a turn. It also writes three title options, the description with full credits, and the thumbnail headline.
• Voices it. ElevenLabs reads the script as Eric, a stock narrator kept permanently so the voice is part of the brand. Any name Eric has never said is voiced on its own and sent to me for approval before it can air.
• Cuts and credits the clips. A small helper server pulls the exact seconds from a match or interview, normalizes the audio, and the on-screen credit plus the source list in the description are generated with the script.
• Renders it. Creatomate draws the graphics kit (roster cards, standings, stat callouts, quote captions, end screen) and the 1080p video. Before a credit is spent, the assembler proves no line runs past its slide.
• Hands it to me. Video, thumbnail and description land in Drive on Monday night. I review, upload and schedule for Tuesday. A Short, when news breaks, lands the same way with a ready-to-paste post for YouTube and X.

How the agent works

The weekly episode, left to right. Every box is an n8n workflow, a Data Table, or a script the workflow calls on the helper server.

A Short takes a shorter road. The credit guard is the box that matters: it keeps enough render budget in reserve for every remaining Tuesday, so a busy news week can never cost the episode.

The stack

• n8n Cloud runs the workflows: sweep, Short builder, voice, render service, Drive delivery, health check. Data Tables hold the story inbox, the seen-items window, the storyline memory and the credit ledger.
• Claude does the thinking: a fast model triages the sweep, a stronger one writes scripts and Shorts. Both load the playbook rules on every run.
• ElevenLabs for the narrator, Creatomate for graphics and video, Google Drive as the handoff point.
• A $6 DigitalOcean box does what n8n Cloud can't: pulls clips and transcripts with yt-dlp, normalizes audio with ffmpeg, assembles the render script, and hosts files at unguessable URLs for seven days.

How I built it

1. Rules before code. I wrote the editor rules and the footage policy first, from YouTube's own monetization criteria, then built the pipeline to enforce them. The script writer never sees a clip it isn't allowed to use.
2. Prove the hard parts first. Transcripts failed with HTTP 429 from a server until the helper used cookies from a throwaway account. Auto-captions garble gamertags (Scrap became "Scump"), so the quote miner cross-checks every name against a roster glossary before a quote can be used.
3. Timing is a check, not a hope. An early mock cut Eric off mid-sentence. The fix was structural: voice durations come from the audio files, every slide holds a second past the last word, and the build fails if a line overruns. A clipped word is a blocking defect.
4. A pronunciation gate. Auto-guesses got player names wrong (Huke as "Hook"). Every new name is voiced as a standalone sound bite and waits for my approval in an approved-names file. Once approved, it never gets asked again.
5. Credits as a budget. A ledger table tracks every render. The Short builder checks it first and skips a Short rather than spend what Tuesday needs. Review renders run at 480p; only the approved cut renders at 1080p.
6. Production reality. A full set of clips expired mid-review and the render failed, so helper files now live seven days instead of three, and a preflight script checks every linked file is live before a credit is spent.
7. Episode 1, then the story rules. The first episode shipped on time and was accurate, and my own review said "information heavy." That review became story-rules: protagonist and turn, clip cadence, one planted callback. Episode 2 runs on them.