gtapps/weekly-digest

Weekly Digest

Produce a weekly synthesis from the archived brief notes. All data comes from .claude-code-hermit/briefs/ — no re-fetching of sources. Cheap to run.

Steps

1. Read the week's briefs. Glob .claude-code-hermit/briefs/*.md (exclude the weekly/ subdirectory), then filter to files whose filename date prefix falls within the past 7 days (today minus 6 days through today). Read each matching file (frontmatter + body).

   • Directory absent or empty → deliver "No briefs archived this week." and stop.
   • Fewer than 3 briefs → note the gap, produce the digest from whatever is available.
   • Track the actual count of files read — this is brief_count for the archive.

2. Synthesize content. From the brief bodies:

   • Top stories — items appearing across multiple briefs or escalating in coverage. Cite which day(s) they appeared.
   • Emerging themes — topics absent early in the week that dominated late in the week.
   • Faded stories — topics in ≥2 consecutive early-week briefs that appear in no late-week brief. Surface at most 2; omit the section if none meet the threshold.
   • Category activity — count items per category directly from brief bodies across all briefs. Do NOT use top_categories frontmatter for this tally — it is pre-truncated to top 5 per brief and undercounts categories that consistently rank below the cutoff.

3. Source performance. Read feed-sources.md for the full source list, then from frontmatter across the week:

   • Reliable sources — appeared in sources_used in ≥5 briefs.
   • Unreliable sources — appeared in sources_skipped more often than sources_used. Include the skip count (e.g. "Source X (skipped 8/12 briefs)").
   • Silent sources — in feed-sources.md but in neither sources_used nor sources_skipped across any brief this week (may never have been attempted). Keep to 3–5 bullets, actionable framing.
   • sources_skipped (fetch failures) is the reliability signal. sources_quiet (0 items on a quiet day) is NOT — never count quiet days against a source's reliability.

4. Write the digest. Format:

   📅 Week of [Mon DD] – [Sun DD, YYYY]
   [N] briefs | [total items across week]
   
   **This week's story**
   [1–2 sentences on the dominant narrative of the week]
   
   **Top stories**
   - [Story] — [days] — [one line on why it matters]
   
   **Themes**
   - Emerging: [theme]
   - Faded: [topic]   ← omit if none met the threshold
   
   **Categories this week**
   [Cat A: 18 items · Cat B: 11 · Cat C: 7 · ...]
   
   **Source notes**
   - Reliable: [source], [source]
   - Skipped often: [source] (N/14 briefs) — consider removing or fixing
   - Silent: [source] — never fetched, check if still relevant
   

   Tone: slightly more reflective than an evening brief. One paragraph of narrative before the lists is fine.

5. Deliver via the Operator Notification protocol in CLAUDE.md § Operator Notification (core resolves the channel and falls back to push when no channel is reachable). text is the digest from step 4.

   • For the push-fallback branch, condense to a single line (per § Operator Notification push format): lead with the week's dominant theme, then the brief count. Example: Quiet week on AI infra, 12 briefs — open CC to read.
   • On resolve miss or send failure, write the digest to .claude-code-hermit/compiled/pending-delivery.md (frontmatter title, type: pending-delivery, created, brief_path pointing at the weekly archive path). Do not retry. This queue supersedes the protocol's SHELL.md-logging branch — don't also log the digest to SHELL.md Findings or record a channel-send-unavailable issue.

6. Archive to .claude-code-hermit/briefs/weekly/YYYY-WNN.md (create the directory if absent). Use the ISO week number (e.g. 2026-W15). Use actual counts from steps 1–3, not placeholders:

   ---
   date: YYYY-MM-DD        # Sunday of the week
   week: YYYY-WNN
   title: "Weekly Digest — YYYY-WNN"
   created: YYYY-MM-DDTHH:MM:SS+HH:MM  # ISO 8601 with timezone
   tags: [brief]
   brief_count: N       # actual number of brief files read in step 1
   top_categories: [Cat A, Cat B, Cat C]   # top 3 from the full body tally in step 2
   sources_used: [Source A, Source B]      # union of all sources_used across the week
   sources_skipped: [Source C]             # union of all sources_skipped across the week
   ---
   

7. Reaction-feedback aggregation. Glob .claude-code-hermit/compiled/brief-feedback-*.md. If no such files exist, SKIP this step silently — the reaction-feedback producer is a channel-layer concern not shipped with this plugin, so its absence is expected, not an error. When logs do exist:

   • Read the two most recent files (current + previous month if both exist).
   • Parse each line (format: YYYY-MM-DD <slot> | <emoji> | source: <name> | topic: <topic>).
   • Filter to entries within the past 7 days (the week window).
   • Count 👍 and 👎 per source. If any source has ≥3 reactions total in the window, add a brief note to the digest's Source notes section (e.g. "Source A — 5 👍 this week", "Source B — 3 👎 this week (consider deprioritizing)").
   • Persistent negative signal: if a source's 👎 count exceeds its 👍 by 3 or more over the month window (both files), create a proposal via /claude-code-hermit:proposal-create to deprioritize or remove that source (Evidence Source: scheduled-check/source-health).

Notes

• Reads from the archive, never re-fetches. If a brief was skipped that day, it simply won't appear — note the gap rather than backfilling.