Magic Agency's AI UGC bot had a deceptively simple brief: request a character-based image or video in Discord and receive a useful asset. Production reality was harder. RECAST treated access, provider execution and creative approval as separate layers, then proved the technical generation path without declaring the workflow fully production-ready.
Evidence at publication: the live bot was active, post-restore task activity existed, provider balance/status probes worked and a controlled Seedance sample completed for four profiles. Visual quality approval and broad operator readiness remain open, so this case is a pilot.
The company
Igor's Magic Agency project centers on Discord-based AI UGC production. Operators should be able to submit an image or video request, select the appropriate character or model profile and receive generated media through controlled provider paths.
That workflow sits at the intersection of permissions, secrets, prompts, reference packs, task state, model behavior and human taste. A healthy service process proves only one of those layers.
The constraint
The project carried several overlapping risks: locked channels could prevent real operators from submitting work; historical provider failures could leave stale tasks; weak reference metadata could degrade consistency; provider switching could obscure the actual cause; and a technically completed render could still fail visual review.
These issues became confused during a “down all day” report on 27 June. Treating the incident as one generic outage would have encouraged changes to the wrong layer.
Why existing tools were not enough
Generation providers can return media, and Discord can accept commands, but neither establishes that the correct operator had access, the right profile was selected, the intended prompt version ran or the final asset met the agency's standard.
The missing production layer needed traceability: permission diagnosis, provider probes, task ledgers, reference metadata, backups, prompt/version discipline and human sample review. Without it, switching models may change the symptom while preserving the failure.
Transformation map
Before
Discord request -> uncertain permission/provider/prompt state -> render or stale task -> unclear diagnosis -> repeated model changes
After
Authorized channel -> profile + reference metadata -> controlled image/video route -> task ledger -> technical sample -> Daniel/Igor visual QA -> production decision
Technical success and creative approval are two different gates in this route.
One workflow, end to end
- Access: An approved content operator enters the designated AI UGC channel with the required Discord permissions.
- Request: The operator submits a simple image or video brief against a named character/model profile.
- Preflight: The workflow validates provider constraints, profile references and required metadata before creating the task.
- Generation: The request travels through the selected GPT Image, Kie, Seedance or controlled video path.
- Ledger: Task status, provider response and resulting artifact remain traceable rather than disappearing inside a bot message.
- Technical QA: The team confirms that the request completed and the output can be retrieved.
- Visual QA: Daniel or Igor judges identity, motion, composition and brand usefulness before the route is approved for broader production.
The source proves the first six steps for a controlled four-profile Seedance sample. It does not retain final broad creative approval.
What changed
- Diagnosed: The June 27 availability complaint was best explained by channel access after AI UGC permissions were restricted.
- Separated: Discord access, provider reliability and creative quality became distinct investigation lanes.
- Protected: Backups and a safe-recovery path were captured before production mutation.
- Inspected: Live service state and task ledgers were audited; provider balance/status probes returned successfully.
- Tested: A final Seedance technical sample package completed across four profiles.
- Still in pilot: Visual QA, reference metadata, prompt/version hardening, stale submitted-row repair and controlled simple image/video QA remain.
- Financial result reported by client: None retained.
Proof
- Service receipt: The VPS-hosted Discord bot was active during the live audit.
- Activity receipt: Post-restore task activity existed in the retained operating record.
- Provider receipt: Kie balance and status probes worked.
- Sample receipt: Four profiles completed the final controlled Seedance technical package.
- Investigation receipt: The record includes a Discord-permission conclusion, provider-failure classification, prompt/video-model hypotheses and Kling preflight constraints.
- Boundary: A completed technical sample is not a claim that every output met brand standards or that broad production was approved.
Technology - revealed last
The pilot runs as a VPS-hosted Discord bot under systemd, with service code at /opt/igor-ai-ugc-bot and runtime state under /var/lib/igor-ai-ugc-bot. It uses the Creative Bot account, Magic Agency guild routes, Kie provider probes, GPT Image/OpenAI image paths, Seedance video paths, Kling preflight rules, task ledgers, reference packs and backup baselines. Local rollback documentation preserves the safe recovery route.
Client quote
No source-backed verbatim client quote is published. The outage wording is incident context, not a testimonial, and is not repurposed as one.
Next transformation
See Yannick Schatz's live agency operating layer for a UGC engine that sits inside a broader, actively used departmental system.
Get your AI blueprint
If a creative bot is “running” but the team cannot prove access, version, output and approval state, the missing product is the production workflow around the model. Book a free AI audit to map it.
Source record
OPENCLAW-DEEP-CLIENT-CASE-STUDIES.pdf, page 24.RECAST_CASE_STUDIES_2026_UPDATED.pdf, page 20, documentRA-CASES-001, version 2.0.RECAST_WEBSITE_MASTER_BLUEPRINT.md, sections 10.5, 11 and 58.