Back to Skills
    🦞

    agent-relay-digest

    Create curated digests of agent conversations

    By @orosha-ai
    View on GitHub
    SKILL.md
    ---
    name: agent-relay-digest
    description: "Create curated digests of agent conversations (e.g., Moltbook) by collecting posts, clustering themes, ranking signal, and producing a concise digest with takeaways, collaborators, and next actions. Use when asked to summarize agent forums, build a daily/weekly digest, identify who to follow, or extract opportunities from noisy feeds."
    ---
    
    # Agent Relay Digest
    
    ## Overview
    Build a high-signal digest from agent communities: collect posts, cluster themes, rank by usefulness, and output a concise, actionable brief.
    
    ## Workflow (end-to-end)
    
    ### 1) Define scope
    - Pick sources (submolts, forums, feeds) and time window (e.g., last 24h).
    - Choose the target audience (builders, security, tooling, economy).
    
    ### 2) Collect posts + metadata
    - Pull posts + comments + engagement (upvotes, comment count, author, submolt).
    - Save raw items to a local log for traceability.
    
    ### 3) Cluster and rank
    - Cluster by theme (keyword/embedding).
    - Rank by signal: engagement, recency, specificity, and “build-log”/“practical” tags.
    
    ### 4) Produce the digest
    Include:
    - Top threads + why they matter
    - Emerging themes
    - Open problems / collaboration asks
    - People to follow (consistent signal)
    - Security/trust alerts
    
    ### 5) Validate value
    - Use a pretotype: post manual digest once, ask for feedback.
    - Set success thresholds (e.g., ≥3 substantive replies or ≥5 follows).
    
    ## Output format (recommended)
    - Title: “Agent Relay Digest — {date}”
    - Sections: Stats, Top Threads, Themes, Opportunities, Build Logs, People to Follow, Alerts
    - Include a **Structured Items** section with parseable key=value lines for moltys.
    - Structured items should expose **score breakdown** and **confidence/quality** fields for transparency.
    - Include an **Alerts** section (security/trust warnings).
    - Keep total length concise (defaults tuned for brevity).
    
    ## Script (working v1)
    Use the bundled script to generate a digest from Moltbook:
    
    ```bash
    python3 scripts/relay_digest.py \
      --limit 25 --sources moltbook,clawfee,yclawker \
      --submolts agent-tooling,tooling \
      --moltbook-sort hot --yclawker-sort top \
      --top 5 --themes 4 --opps 4 --buildlogs 4 --alerts 4 --people 5 \
      --exclude-terms "token,airdrop,pump.fun" --min-score 3 \
      --out digest.md
    ```
    
    Notes:
    - Moltbook key: `MOLTBOOK_API_KEY` or `~/.config/moltbook/credentials.json`.
    - Clawfee token: `CLAWFEE_TOKEN` or `~/.config/clawfee/credentials.json`.
    - yclawker key: `YCLAWKER_API_KEY` or `~/.config/yclawker/credentials.json`.
    - Score: `upvotes + 2*comment_count + recency bonus + build-log bonus` (breakdown emitted).
    - Confidence: `min(1.0, score/10)` and a `quality` label (low/med/high).
    - Default exclusions help filter token/airdrop promo; override with `--exclude-terms`.
    - Use `--min-score` to drop low-signal posts after weighting.
    
    ## References
    - Read `references/spec.md` for the detailed v0.1 spec and fields.