Monitored-source editorial pipeline
Build the source workflow in its stated order. Keep final editorial judgment manual.
Source workflow
Set up the Hermes board and roles
Install Hermes from the Hermes site and use its model picker. Create five profiles:
- scout-x monitors X for content opportunities;
- scout-web monitors newsletters, Reddit, and YouTube;
- orchestrator groups signals and scores content potential;
- researcher verifies evidence and audits existing coverage;
- ideator turns researched ideas into angles.
Create a dedicated Hermes kanban board named content-engine. Keep content work separate from coding tasks and reminders. Use triage for new signals, todo for promising ideas, ready and running for active work, blocked for human review, and done for history. Hermes stores the board in local SQLite so later agents retain sources, scores, comments, and the last completed step.
Set the source's starting dispatcher values: 60-second dispatch interval, maximum two jobs in progress, failure limit two, and automatic child promotion on. Set `FIRECRAWL_API_KEY` when clean web-to-Markdown collection is required. Treat the maximum of two as the source's spend-control starting point.
Curate and schedule the scouts
Describe the audience, subject, and content goals. Research the people and publications shaping that conversation, review the list manually, remove repeaters, and add trusted sources. Start with 20 sources, inspect the results, then grow toward 50.
Schedule both scouts for 6am and keep the 20-card cap until the morning board is comfortable to review. Each signal card records what happened, why the audience may care, where it came from, and its topic.
Cluster and score signals
Have the orchestrator group related signals, remove repeats, and score the content opportunity rather than the excitement of the source:
- 70 or higher earns research work;
- 50-69 stays in todo with the missing evidence or angle recorded;
- below 50 is archived with a short reason.
Gate ideation on two research tasks
For each score of 70 or higher, create two parent tasks: one evidence task for verified facts, dates, numbers, quotes, and primary links; and one coverage task for the strongest existing pieces, angles they own, and useful unanswered questions. Create the ideator's package-angles card with both research cards as parents.
Send thin evidence back to hold. When existing coverage owns the proposed angle, sharpen the contribution or archive it. Let the ideator see only ideas with verified evidence and a defensible opening, then prepare three angles with the sources, evidence, and coverage gap.
Keep the final decision manual
At 7am, run `/kanban list` in Telegram and open candidates with `/kanban show`. Review the three angles, source list, evidence, and missed-coverage question. Approve the angle still worth making at zero engagement. Reject the rest with a specific reason. Keep the approved card open while drafting, follow its source links, and change the angle when the editor's reading supports a better one.
Tune in small weekly batches
Add the optional tuner profile to review outcomes and propose improvements. Run it every Friday. Approve useful edits manually and limit each cycle to three changes. Review model spend before increasing concurrent work.
Use the source's 30-day review for signal-to-approval rate, approval-to-publish rate, cost per published idea, and average time from signal to decision.
XSkills operational adaptations
Require the audience, subject, content goals, reviewed starting watchlist, and Telegram destination before execution. If any is missing, ask for it and stop before claiming setup.
Before changing Hermes, verify that the current installation exposes the hermes profile create command, kanban initialization and board selection, dispatcher configuration, scheduling, Telegram, and the cited environment configuration. If a command or integration is unavailable, return the exact board, role, dispatcher, task-dependency, and schedule specification without claiming it was configured. Do not invent a replacement command.
When the source's embedded prompt text is unavailable, create executable prompts from only the stated role contracts: scouts return at most 20 four-field signal cards; the orchestrator clusters, deduplicates, and applies the three score bands; the researcher returns the two named packs; and the ideator returns three angles only after both parent tasks complete. Label these prompts as XSkills operational restatements.
XSkills safety adaptations
Use only sources and accounts the operator is allowed to access. Keep new integrations, spend increases, workflow-policy changes, and publishing behind explicit human approval. Respect privacy, copyright, licensing, platform terms, credentials, and collection limits.
Return
For setup-only work, return the configured components or fallback specification and stop. After a scout run, also return the morning shortlist. Include human decisions and approved weekly changes only after those events exist; otherwise mark them not yet available.
Source attribution
Independently compiled by XSkills from VibeMarketer's X Article about building an autonomous content engine with Hermes. VibeMarketer did not author or review this skill. Current-capability fallback and the authorization, privacy, rights, platform, spend, and publishing boundaries are XSkills adaptations. Attribution does not imply endorsement.