news-personal-digest

name

news-personal-digest

description

Use when generating the personalised daily news digest at delivery time. Reads today's master doc + the interests profile + the runner-prefilled `## Recent feedback (last 7 days)` block; never re-rates items. Picks 4-12 items grouped by category, writes 2-4 sentence briefings with a "Why you care" line per item, and returns a structured JSON block for the daemon to render to Telegram with rating buttons. Idempotent — safe to re-run for the same date. Use whenever the daemon's runner instructs you to produce or refresh a personalised news digest.

News Personal Digest

You produce one personalised daily digest from a categorised master document, an interests profile, and a recent-ratings preference block. The daemon delivers your output as inline-keyboard Telegram messages with 👍 / 👎 / ⭐ buttons per item; tomorrow's run will see those ratings as preference signal, so what you choose today shapes what gets surfaced tomorrow.

Your inputs (already prepared by the runner):

• Target date — the digest is for this date's master doc.
• Master doc at 01-Projects/News/daily/<target_date>-master.md — the candidate pool, already categorised with entity wikilinks.
• Interests profile at 01-Projects/News/interests-profile.md — Adrian's curated context.
• Recent feedback block — already inlined in the prompt under ## Recent feedback (last 7 days), grouped by ⭐ / 👍 / 👎.

Why these constraints matter

• The feedback block is signal, not items. Don't re-rate, don't include them in today's output, don't second-guess them. Their job is to tell you what kind of items resonated recently so you can lean into similar topics (and away from rated-down ones) when picking from today's candidates.
• "Never surface" is a hard filter. The interests profile lists topics Adrian has explicitly opted out of seeing. Drop matching candidates entirely — don't reword to dodge the filter, don't include "for context".
• Briefings beat headlines. The daemon delivers your briefing text as the message body, not a link. Adrian reads it, then decides whether to follow the link. So 2-4 sentences with concrete content (numbers, version strings, named people, mechanism) — not a paraphrase of the title.
• "Why you care" must be specific. Generic relevance ("This is interesting because AI") wastes the line. Anchor each one in something concrete from the profile — a project, a personal context detail, a recent rating pattern. If you can't write a specific reason, the item probably doesn't belong in today's digest.

Procedure

1. Read the interests profile

vault_read("01-Projects/News/interests-profile.md")

The profile has four sections you'll use:

• Active interests — wikilinks to project _index.md files. Anything adjacent is high signal.
• Personal context — life-stage / financial / relational facts that feed "Why you care" lines.
• Always surface — named entities and topics that should be highlighted whenever they appear.
• Never surface — hard filters (drop matching candidates).

If you need richer grounding for the "Why you care" lines, optionally vault_read 1-2 of the project _index files linked under "Active interests". Don't read more than that — token budget matters and the index files are dense.

2. Read the master doc

vault_read("01-Projects/News/daily/<target_date>-master.md")

Every bullet under a ## <Category> heading is a candidate. The bullet's [[00-Inbox/news/<date>/<slug>|source]] wikilink is the source-note path you'll use as source_path in the output JSON.

3. Treat the feedback block as preference signal

The runner has prefilled ## Recent feedback (last 7 days) in the prompt above. Read it once. Build a mental model of which topics, entities, and categories Adrian has rated:

• ⭐ items — strongest positive signal (lean toward similar topics; weight at ~3× a 👍).
• 👍 items — positive signal.
• 👎 items — negative signal (push similar topics down, but don't blanket-block — a topic with one 👎 isn't dead).

You will NOT re-rate any of these items in today's output. They are tomorrow's history; your job is today's selection.

4. Apply hard filters

Drop every candidate whose title or summary matches anything in the profile's "Never surface" list. Be liberal — if a candidate is even close to a never-surface topic (e.g. "fintech that mentions crypto in passing" against a "Crypto" filter), drop it. The user opted out for a reason.

5. Score and pick 4-12 items

For each remaining candidate, score relevance to:

• (i) Active interests + Always surface entries (high weight).
• (ii) Recent ⭐ / 👍 patterns (medium weight).
• (iii) Topics matching recent 👎 (negative weight).

Pick the top 4-12. Aim for 6-10 — that's the readable sweet spot. Diversify across categories: don't stuff one category with seven items just because they all score well. Two strong items in five categories beats seven items in one.

6. Read source notes for the picked items only

For each chosen item:

vault_read("<source_path from master-doc wikilink>")

The source note's body is what grounds your briefing. Pull the concrete details — dollar amounts, version numbers, named people, dates, mechanism — that make the briefing more than a paraphrase of the headline.

Deep-dive enrichment

Some master-doc items have a ### 🔬 Deep dive block beneath them (added by the news-research agent earlier in the chain). When an item you include in the digest has one:

• Fold its synthesis into your "Why you care" line — the deep dive is the sharpest available take; use it.
• If the deep dive has notable "Key facts" or "Sources", surface the single most relevant one inline. Don't dump the whole block; the digest stays skimmable.
• Items without a deep-dive block are briefed exactly as before. Absence is normal — research only covers a few items per day.

7. Group by category, order categories

Group your picked items by their master-doc category (the ## <Category> heading the candidate lived under). Order categories by total signal strength: most-resonant first, so Adrian sees the most-relevant block at the top of the digest. Use the master doc's existing emoji conventions where present — if a category doesn't have an obvious emoji yet, the daemon falls back to "•".

8. Verification (mandatory before returning)

Run this checklist. If any fails, fix it before emitting the JSON, OR emit {"success": false, "error": "<reason>"} and stop.

• (a) Every item has source_path, title, url (or null), briefing ≥80 chars, why_you_care.
• (b) Total items across all categories is in the range [4, 12].
• (c) No item title or briefing matches a "Never surface" pattern from the profile.
• (d) Every source_path appears as a wikilink in the master doc you read in step 2.

9. Emit the structured summary

End your response with a single fenced JSON block matching this schema:

{
  "success": true,
  "rating_signal_summary": "1-line description of how recent ratings shaped this digest",
  "categories": [
    {
      "name": "AI",
      "emoji": "🤖",
      "items": [
        {
          "source_path": "00-Inbox/news/2026-05-09/anthropic-update.md",
          "title": "Anthropic ships Claude 4.7",
          "url": "https://anthropic.com/news/claude-4-7",
          "briefing": "Larger context window (1M default) and 30% faster cache hits. Released alongside a Sonnet 4.6 deprecation timeline targeting Q3.",
          "why_you_care": "Daily-driver model for Automation; agent-infra is built around its specific behaviours."
        }
      ]
    }
  ]
}

Field semantics:

• success: true only if verification passed.
• rating_signal_summary: one sentence describing the shape of recent feedback's effect (e.g. "Boosted AI / Anthropic items based on three recent ⭐; deprioritised generic Series B funding given two recent 👎."). Empty-ish runs (no recent ratings) get something like "No prior ratings — neutral run."
• categories: ordered list. Categories with zero items must NOT appear.
• error: only on success: false. One sentence explaining what failed.

The runner extracts the LAST fenced json block in your output, so progress JSONs earlier in the message are fine — only the last one is parsed.

Quality bar

• Briefings reference content from the source note. "Anthropic shipped a new Claude" is a paraphrase. "1M default context, 30% faster cache hits, Sonnet 4.6 deprecation in Q3" earns the line.
• "Why you care" is anchored. Tie each one to either a named project (link or topic), a personal-context detail (residency, EUR exposure, etc.), or a recent rating pattern. Generic "this matters because AI" is a smell.
• Diversity beats stuffing. A digest with one item in five categories beats seven items in one — variety preserves the daily-routine value.
• Concrete > clever. No "the AI world saw a major shake-up today" — name the entities, the numbers, the mechanism.

Worked example: a good item

{
  "source_path": "00-Inbox/news/2026-05-09/ecb-rate-decision-ab12.md",
  "title": "ECB holds at 3.25%, signals June cut",
  "url": "https://ft.com/...",
  "briefing": "ECB held the deposit rate at 3.25% in line with consensus but Lagarde's press conference language shifted from 'data-dependent' to 'increasingly confident', priced as a 70% probability of a 25bp cut in June. Bund yields fell 8bp on the day.",
  "why_you_care": "Eurozone rate trajectory directly affects personal financial planning; a June cut shifts the timeline for refinancing decisions."
}

Worked example: a verification failure response

If verification fails (e.g., only 3 items survived all filters and no realistic sixth could be reached without violating a hard filter):

{
  "success": false,
  "error": "Only 3 items survived hard filters and ranking; minimum digest size is 4. Master doc had 8 candidates, 4 dropped by 'Never surface' filter, 1 too off-topic."
}

The runner records this as failed_verification and surfaces it to Adrian. Don't pad with weak items to hit the minimum — surface the gap honestly.