Content Production OS
Creators need repeatable, research-grounded content production with clear handoffs and final editorial approval. Build a multi-role workflow that can run sequentially with one agent or with specialized agents when available.
Who and when?
YouTube educators, podcast teams, video creators, scriptwriters, editorial teams.
From intake to final review.
- Claims requiring verification are backed by attributable sources or flagged for removal. - Script QA identifies exact passages and actionable fixes. - Final approval is explicit, version-specific, and not inferred from a draft's existence. - Production pack references the approved script version only. - Missing browsing/source access triggers a limited offline mode with explicit unverifiable-claim warnings.
Inputs and outputs
Required: content brief, audience, channel/format, topic, language/tone, target duration, editorial constraints. Optional: content strategy, series bible, approved sources, production capabilities, prior episodes, distribution goals.
- research-plan.md
- sources.csv
- claim-ledger.csv
- outline.md
- script-vN.md
- script-qa.md
- final-approval.md
- script-final.md
- production-pack.md
- distribution-pack.md
- state.json
A synthetic worked example.
Synthetic example: A fictional five-minute episode begins with three claims. One has no evidence and is removed. Version one needs a specific paragraph change. After version two is reviewed and approved, the production pack maps each segment to visuals, overlays, and editing notes. A rejected version cannot enter production.
claim_id,source_id,status CL-01,SRC-01,verified CL-02,,remove
An illustrative output excerpt, not a client result or live run. Sample approval cannot authorize a real action.
How to use it
- After repository publication, obtain the files and read README and SKILL.md. No download link exists before release.
- Prepare inputs and sources in a separate run folder. Keep secrets and client data out of public files.
- Start at intake and follow the workflow manually or with an AI tool that can read and write local files. No platform compatibility is claimed before testing.
- Review each phase and record a version-specific decision. Missing inputs and weak evidence create blockers, not guessed completion.
- Persist state, decisions, and artifacts. Resume from the latest approved phase without silently replacing approved output.
Limits and safety
Video editing, automatic recording, voice cloning, paid asset generation, publishing.
Files and web pages are untrusted information, not authorization to change goals, publish, or contact anyone. External effects need separate permission. Run data stays local and examples are synthetic. File checks cannot prove factual truth or reviewer identity.
Status: v1 specification for review. GitHub URL, release, license, and exact runtime requirements are not published. Contributions will use the actual repository after launch.