name: cron-prompt description: Rules for authoring high-fidelity, multi-phase cron prompts for Hermes Agent. Ensures consistent pipelining, role assignment, and reporting.

Cron Prompt Engineering

Use this skill when designing or updating cron job prompts to ensure they follow the "Phase Pipeline" architecture for reliability and transparent reporting.

Required Prompt Structure

Every cron prompt MUST explicitly define the following sections:

1. Phase-Based Pipeline

Decompose the task into discrete phases (e.g., Research, Analysis, Writing, Review). For each phase, specify: * Assignee: MUST use the exact profile/department name (e.g., r-and-d, grill-qa, writers, it-ai, str-mkt). Do NOT use generic titles like "Researcher" or "Analyst". * Personality & Skills: Explicitly state which personality (e.g., /personality public-policy) and which skills (e.g., social-listening, research-critism) the assignee must load or apply during this phase. * Job Description: Precise instructions on what to fetch, filter, or generate. * Success Criteria: What defines a completed phase.

2. Telegram Reporting Rules

Define exactly what the user sees in the Telegram thread for each phase: * Minor Phases: Update with a single, concise status line (e.g., "Phase 1: Data collection complete"). * Major Deliverables: If a phase generates a file (e.g., .md report, .csv data): * Provide a Brief (concise summary of findings). * Provide a Tailscale Tunnel URL (formatted as https://vmi3427693.tail8c1aaf.ts.net/files/<full_path>) so the file is clickable. * Final Output: Clear delivery of the digest or primary artifact.

3. Execution Metadata

The prompt should explicitly state: * Workdir: The absolute path where the job operates (e.g., /opt/ai-os/products/ceo/...). * Time Window: The lookback period for data (e.g., "Last 72 hours"). * Recurrence Logic: How often it runs (matching the cron schedule). * Silence Rule: Use [SILENT] if no new high-signal data is found to avoid spam.

Example Blueprint

[PHASE 1: RESEARCHER]
- Assignee: r-and-d
- Skills: topic-news-digest-social-listening
- Task: Scrape [Topic] from [Sources].
- Window: Last 72 hours.
- Telegram: "Phase 1: Scanned [X] sources, found [Y] matches."

[PHASE 2: ANALYST]
- Assignee: grill-qa
- Personality: public-policy
- Task: Filter for relevance and score risks.
- Output: Save detailed analysis to [Path]/[Topic_Slug]_[YYYY-MM-DD]_analysis.md
- Telegram: Brief of findings + https://vmi3427693.tail8c1aaf.ts.net/files/[Path]/[Topic_Slug]_[YYYY-MM-DD]_analysis.md

[PHASE 3: WRITER/REVIEWER]
- Assignee: writers
- Skills: research-critism
- Task: Draft and verify the final digest.
- Telegram: The final high-signal message in thread [ID].

Anti-Patterns to Avoid

• Monolithic Prompts: Avoid one big block of text; AI misses steps without phase boundaries.
• Hidden Paths: Never assume the agent knows where to save; always specify the workdir.
• Broken Links: Always use the Tailscale Funnel URL for files intended for human review.
• Static Filenames: Never use generic or static filenames (e.g., analysis.md) that will overwrite previous runs or conflict with other projects. Always append timestamps, dates, or unique project slugs.