newsletter

name	newsletter
description	Generate today's daily news report at reports/newsletter/YYYY-MM-DD.md by performing WebSearch + WebFetch over the user's 3-layer context (Identity / Attention / Impulse) and producing Alert + Deep Dive + Discovery sections. Fully automated. Use when the user asks for the daily research newsletter.
gui	{"label":"/newsletter","hint":"Generate daily research report","match":["reports/**/*.md"],"arg":"none","order":70,"mode":"auto"}

/newsletter — Daily News

Conduct all conversation with the user — and write all generated output — in the language defined by .claude/rules/personal-language.md (or the user's input language if that file is absent). Follow the language rules in full — exceptions and translation quality are defined in the Language Rules of .claude/rules/rill-core.md and the vault's personal-*.md overrides, never restated per skill. The English instructions below are for skill clarity, not for output style.

Tool references in this skill (shell, Read, Edit, Glob, WebSearch, WebFetch) describe intent, not Claude-specific tool calls. Each harness should map them to its native equivalent — Claude Code uses its built-in tools as named; Codex CLI uses shell / apply_patch / --search / its own URL fetcher as appropriate.

Performs web searches based on the user's Interest Profile, Project entities, workspaces, and journals (3-layer context), generating a daily news report focused on discovering unknowns + alerts. Fully automated (no interaction).

Arguments

$ARGUMENTS — one of the following:

• YYYY-MM-DD (e.g., 2026-03-11) → Generate report for the specified date
• Omitted → Generate report for today

Procedure

Phase 0: Initialization + Deduplication Data Collection

1. If argument is YYYY-MM-DD: Use that date as target
2. If argument omitted: Use today's date as target
3. If reports/newsletter/YYYY-MM-DD.md already exists: Overwrite without confirmation (version history in Git)
4. If reports/newsletter/ directory doesn't exist, create it

3-Layer Context Collection (execute in parallel):

5. Identity: Interest Profile — Read knowledge/self/profile.md + knowledge/self/interests.md + knowledge/self/direction.md (the Identity layer is the static + medium-velocity combination)
6. Identity: Project Entities — Read all projects/*/_project.md (get Competitors, Watch Keywords). (ADR-080: projects moved from knowledge/projects/*.md flat layout to top-level projects/{slug}/_project.md per-directory layout)
7. Attention: Workspaces — From all directories under workspace/, read _workspace.md (or _session.md) for workspaces created within the past 2 weeks, extracting (regardless of status — both active and completed):
   • # heading (workspace name)
   • Description text immediately after heading (1-2 lines)
   • Directory name (date + topic slug)
   • ※ Don't depend on tags or name: field as they can be unreliable. Use heading + description
8. Impulse: Journals — Read full text of past 7 days of files from inbox/journal/ (~100 lines). Prefer _organized/ version if same-named file exists. From journals, extract only project/career/technology/business-related topics as interest signals (ignore shopping, daily logs, etc.). Do not include emotions or personal information in newsletter body

Deduplication Data Collection (execute in parallel with above):

9. Past 2 weeks of newsletters — Read past 2 weeks from reports/newsletter/, extracting:
   • Used keywords (keywords frontmatter)
   • Referenced URLs (main sources section)
   • Covered topics (section headings)
10. Known knowledge index — Glob knowledge/notes/*.md filename list (for known information filtering)

Phase 1: Search Strategy Construction

Integrate the 3-layer context (Identity / Attention / Impulse) and generate search keywords divided into 3 categories.

3-Layer Context Model:

• Identity (self/profile.md + self/interests.md + self/direction.md + projects/): Stable interest categories. Changes monthly
• Attention (workspace/ past 2 weeks): Currently active themes. Changes weekly
• Impulse (journal/ past 7 days): Not-yet-formed interests/thoughts. Changes daily

Alert Category (2-3 items) — Focused on detecting changes

Source: projects/*/_project.md Competitors + Watch Keywords + self/interests.md Career Purpose: Detect competitor movements and changes directly affecting projects

• Generate keywords from each Active Project's Watch Keywords and major competitor names
• Include Career section monitoring targets (e.g., "Anthropic Japan hiring")
• Obligations: Only keywords for detecting major changes like regulatory/institutional changes
• watch-domestic: true projects must include at least 1 Japanese keyword
• Alert is only for "specific changes affecting project management decisions": Competitor pricing changes, feature releases, fundraising, partnerships, executive moves. Industry-wide trends or technology standard developments go in Discovery, not Alert

Deep Dive Category (1-2 items) — Thorough investigation

Source: self/interests.md Deep Interests + weighted by Attention layer (workspace themes) Purpose: Deeply explore one theme to a level where user thinks "I didn't know that"

• Select 1-2 themes from Deep Interests
• Workspace weighting: Themes being worked on in past 2 weeks' workspaces get higher selection probability (e.g., CRM design workspace exists → CRM × AI latest trends as Deep Dive candidate)
• Journal weighting: Business/tech themes mentioned in past 7 days' journals are also included as candidates (e.g., "want to release Rill" → OSS PKM tool release strategy)
• Forced rotation: Do not select themes overlapping with past 2 weeks' newsletter covered topics
• Evaluate theme's "depth potential": Avoid themes unlikely to yield new info on deeper investigation (e.g., fixed content like invoice system rules)
• Generate 2-3 search keywords per theme (varying angles)

Discovery Category (2-3 items) — Findings & Recommendations

Source: self/interests.md Curiosity + adjacent areas of Attention/Impulse layers Purpose: "If you're interested in this, you might find that interesting too"

• Select themes from Curiosity section
• Workspace adjacent exploration: The "neighbor's neighbor" of recent workspace themes yields highest-value Discoveries
  • e.g., During CRM feature design → conversational CRM, AI-native CRM latest trends
  • e.g., During content marketing automation → AI ghostwriting ethics, synthetic media
• Journal-originated exploration: Include themes appearing in journals but not yet as workspaces
• Adjacent exploration: Consciously seek the "neighbor's neighbor" of Deep Interests and Active Projects
  • e.g., Interested in Claude Code → compiler theory, programming language design
  • e.g., PKM × AI → cognitive science, spaced repetition
• Prioritize topics not yet in knowledge/notes/ filename list
• Serendipity: OK to intentionally include 1 "unexpected" topic

Common Keyword Generation Rules:

• Prefer English keywords (for web search coverage)
• Make phrases specific ("Claude Code agent SDK release 2026" not "AI")
• Include the search year for latest information
• Keywords used in past 2 weeks' newsletters are prohibited
• Record category name (Alert/DeepDive/Discovery) and 1-line selection rationale as internal notes for each keyword

Phase 2: WebSearch Execution

For each keyword generated in Phase 1:

1. Search with WebSearch (each keyword independent, execute in parallel where possible)
2. Collect top 3-5 results from each search
3. Determine publication date where possible (from URL, snippet, article content)
4. Deduplication: Exclude URLs referenced in past 2 weeks' newsletters
5. Aggregate all results into temporary lists by category

Phase 3: Deep Investigation

Fetch full content with WebFetch by category:

Alert Category

• Select only articles where actual change occurred for each competitor/monitoring target (3-5 articles)
• Freshness filter: Only articles published within the past 2 weeks qualify as Alert candidates. Older articles are not accepted as Alerts (may be covered in Deep Dive or Discovery)
• If no changes detected, don't force article retrieval
• Primary source verification (required): Alert candidate facts must be verified by WebFetching primary sources (official sites, official announcements, first-party pages) before inclusion. Do not compose Alerts solely from secondary sources (news articles, blogs). Example: hiring info → WebFetch official careers page, feature release → WebFetch official blog/changelog, fundraising → WebFetch press release. If primary source cannot be WebFetched, explicitly state in Alert body ("could not verify from official page")

Deep Dive Category

• WebFetch top 8-10 most important results for the selected theme
• Prioritize primary information (official announcements, papers, official docs over blog aggregation articles)

Discovery Category

• WebFetch top 3-5 results per theme

Common WebFetch Tasks:

• Publication date confirmation: Determine from in-article dates, URL patterns, bylines
• Deep Dive / Discovery: If WebFetch fails, may substitute with search result snippets
• Alert: Snippet-only substitution prohibited. If primary source WebFetch fails, either explicitly state in Alert body or exclude from Alerts
• Silently skip timeouts and 403 errors (but follow above rules for Alert primary source verification failures)

Selection Criteria:

• Freshness is top priority: Strongly prefer articles from past 1-2 weeks
• Known information filter: Skip content clearly overlapping with knowledge/notes/ filenames
• High relevance to user's projects

Phase 4: Report Generation

Generate at reports/newsletter/YYYY-MM-DD.md with the following structure.

Report Body Language

Write the report body in the language defined by .claude/rules/personal-language.md (English when absent). This skill floods the context with English web sources — translate their vocabulary instead of importing it. A large amount of English in the context is not a reason to mix English words into the report. Translation quality follows the Language Rules in .claude/rules/rill-core.md and the vault's personal-*.md overrides — in particular, never invent literal calques that do not exist in the target language.

Skill-specific exception: the fixed template strings — the section headings and labels defined in the Template below (# YYYY-MM-DD Daily News, ## Alerts, ## Deep Dive:, ## Discovery, ## Research Metadata, and the Alert: / Deep Dive: / Discovery: keyword labels) — stay verbatim English because downstream tooling parses them.

File Creation

First create the file with rill mkfile (to ensure timestamp accuracy):

rill mkfile reports/newsletter --type newsletter \
  --field "keywords=[keyword1, keyword2, ...]" \
  --field "source-count=N" \
  --field "alert-count=N" \
  --field 'deep-dive-topic="Theme name"' \
  --field "discovery-count=N"
# For specific date: add --date YYYY-MM-DD

Then use Edit to append body to the output path (frontmatter is pre-generated by rill mkfile).

Template

# YYYY-MM-DD Daily News

## Alerts

(Generate section only if changes detected. If none, use single line "No notable developments")

> ⚡ [1-line change summary] — YYYY-MM-DD

(Prose with details. Start with time anchor like "On Month Day,".
What changed, why it matters, impact on projects.
Inline source URL + publication date: [Title](URL) (YYYY-MM-DD).
200-400 characters per Alert. Keep Obligations alerts concise.
**Timeline is critical information**: Never be vague about when a change occurred)

---

## Deep Dive: [Theme Name]

> Sources: YYYY/M — YYYY/M (composed primarily from articles in the past N days)
> Why this theme: (1-line explanation of which Interest Profile category and why)

(Prose. Minimum 1000 characters. Write in 3-layer structure:

1. **Facts**: What is happening. Inline source URLs and publication dates
   - Format: [Title](URL) (YYYY-MM)
   - Unknown date: [Title](URL) (date unknown)
2. **Interpretation**: What it means. Technical/business context
3. **Implications**: How it affects user's projects and interests

**Freshness differentiation**: Clearly distinguish between recent developments (past 1-2 weeks)
and background context (older information) at the paragraph level.
Never write old information as if it were new news)

---

## Discovery

### [Theme Name]: [1-line why this is recommended]

(Prose. 300-500 characters. Explicitly state connection points to user's existing interests.
Recommended format: "If you're interested in X, here's why Y is worth watching...")

### [Theme Name]: [1-line why this is recommended]

(Same. 2-3 Discovery sections)

---

## Research Metadata

- **Search Keywords**:
  - Alert: keyword1, keyword2
  - Deep Dive: keyword3, keyword4
  - Discovery: keyword5, keyword6
- **Information Freshness**: YYYY-MM-DD — YYYY-MM-DD
- **Source Count**: N items
- **Interest Profile Reference**: Deep Interests [theme], Curiosity [theme]
- **Key Sources**:
  - [Title](URL) (YYYY-MM)
  - ...

Post-output

Display a summary (3-5 lines) and finish. Do not transition to assistant mode.

Rules

• Fully automated: No interaction. Everything from keyword selection to generation is automatic
• Prose-based: Write engaging prose, not bullet point lists
• All facts require source URL + publication date: No hallucination. Use only information obtained from WebSearch/WebFetch, with source URLs and publication year-month for all facts
• Explicit information freshness: Never write old information as new news. Clearly distinguish "latest developments" from "background information" at the paragraph level in the body
• Convey the unknown: Avoid repeating information the user already knows (topics in knowledge/notes/). Prioritize new information, new perspectives, new connections
• Interest Profile is a guide: Don't be overly constrained by categories. Especially in Discovery, unexpected suggestions are welcome
• Alert only when change occurs: If no change, summarize as "No notable developments." Don't force-create sections
• Deep Dive goes deep: 1000+ characters. Not a superficial summary, but deep enough to make the reader think "I didn't know that"
• Thorough deduplication: Eliminate overlap with past 2 weeks' keywords, URLs, and topics
• Re-execution overwrites: Version history is in Git
• Never modify inbox/ original files (read-only)