ai-daily-digest

分类: 调研与分析 | 上传者: smykla-skalski | 下载: 0 | 版本: v1.0（最新）
Daily AI news digest covering technical advances, business news, and engineering impact. Aggregates from research papers, tech blogs, HN, newsletters. Use daily for staying current on AI developments.
更新日志: Source: GitHub https://github.com/smykla-skalski/sai
目录结构

当前层级: claude/ai-daily-digest/skills/ai-daily-digest/
📁 references/
- 📄 output-template.md 9.7 KB
- 📄 search-patterns.md 12.0 KB
- 📄 sources.md 8.3 KB
📄 SKILL.md 10.6 KB
SKILL.md

---
name: ai-daily-digest
description: Daily AI news digest covering technical advances, business news, and engineering impact. Aggregates from research papers, tech blogs, HN, newsletters. Use daily for staying current on AI developments.
argument-hint: "[--focus technical|business|engineering|leadership|all] [--notion-page-id ID] [--no-notion]"
allowed-tools: AskUserQuestion, Bash, Read, Task, ToolSearch, WebFetch, WebSearch, Write
user-invocable: true
disable-model-invocation: true
---

# AI Daily Digest Skill

Generate comprehensive daily AI news digest with technical, business, and engineering coverage.

## Arguments

Parse from `$ARGUMENTS`:

- `--focus [technical|business|engineering|leadership|all]` — Default: all
- `--notion-page-id [UUID]` — Notion parent page ID for digest publishing (overrides env var)
- `--no-notion` — Skip Notion publishing entirely (archive-only mode)

## Configuration

### Notion Parent Page ID (required for publishing)

Resolve the Notion parent page ID using this precedence (first match wins):

1. `--notion-page-id` argument
2. `NOTION_PARENT_PAGE_ID` environment variable
3. Interactive prompt — use AskUserQuestion at runtime

When prompting the user, provide these instructions for finding the page ID:

- Open the target parent page in the browser
- Copy the URL (`https://www.notion.so/Page-Title-{32-hex-chars}`)
- Extract the last 32 hex characters, insert hyphens as `8-4-4-4-12` for UUID format
- Alternative: click "Share" → "Copy link" and extract the ID

Recommend persisting via env var in `~/.zshrc` / `~/.bashrc`:

```bash
export NOTION_PARENT_PAGE_ID="your-page-id-here"
```

Or in `~/.claude/settings.json` under the `"env"` key:

```json
{
  "env": {
    "NOTION_PARENT_PAGE_ID": "your-page-id-here"
  }
}
```

## Preprocessed context

- Data directory: !`echo "${XDG_DATA_HOME:-$HOME/.local/share}/sai/ai-daily-digest"`
- Today: !`date +%Y-%m-%d`
- Day of week: !`date +%A`
- Notion page ID (env): !`echo "${NOTION_PARENT_PAGE_ID:-}"`

## Persistent Data Directory

All persistent state and generated artifacts are stored in the data directory from preprocessed context above. This path is independent of the plugin cache and project working directory — artifacts survive plugin updates and work from any project.

Use the pre-resolved data directory path for all file operations. Do NOT use `./findings/` or other relative paths — they may resolve to the plugin cache and be lost on updates.

## State Files

All state stored in the persistent data directory.

### Last Run Date (`.last-run`)

Format: `YYYY-MM-DD`. Read on startup to calculate date range. If missing, default to past 7 days.

### Covered Stories (`.covered-stories`)

Pipe-separated: `{date}|{story_id}|{url}` — one story per line.

- `story_id` — Normalized: lowercase, hyphen-separated, key terms (e.g., `falcon-h1r-7b-release`, `xai-20b-funding`)
- Keep last 300 entries (trim oldest when exceeding)
- Prevents duplicate stories across days via story_id fuzzy matching and URL matching

Example:

```text
2026-01-28|deepseek-r1-release|https://api-docs.deepseek.com/news/news250120
2026-01-29|falcon-h1r-7b-release|https://falcon-lm.github.io/blog/falcon-h1r-7b/
```

## Workflow

### Phase 1: Setup

1. Read [references/sources.md](references/sources.md) for source URLs and tiers
2. Read [references/output-template.md](references/output-template.md) for digest format
3. Parse arguments for `--focus` area and `--notion-page-id`
4. **Resolve Notion page ID** — if `--no-notion` is set, set `notion_page_id` to `null` (archive-only mode).
   Otherwise check in order: `--notion-page-id` arg → Notion page ID from preprocessed context → prompt user interactively.
   Store resolved value as `notion_page_id` for Phase 18.
   If the preprocessed value is empty and user declines to provide an ID, skip Notion publishing (archive-only mode).
5. **Set up persistent data directory** — use the data directory from preprocessed context as `DATA_DIR`. Run `mkdir -p "$DATA_DIR"` to ensure it exists.
6. Read `$DATA_DIR/.last-run` — set date range from last run to the today value from preprocessed context
7. Read `$DATA_DIR/.covered-stories` — build in-memory `covered_ids` and `covered_urls` sets
8. Check day of week from preprocessed context — if Friday, enable weekly recap mode (Friday Weekly Recap section in search patterns)

### Phases 2-15: Research

**Before starting Phase 2, read [references/search-patterns.md](references/search-patterns.md) in full for all search queries, source-specific patterns, and the Friday Weekly Recap section.**

Spawn a `general-purpose` research agent for each phase (or batch of independent phases).
Pass each agent: the date range, `covered_ids` and `covered_urls` sets, the focus area, and the relevant section from [references/search-patterns.md](references/search-patterns.md).
Phases 2-5 are independent - spawn them in parallel. Subsequent phases can be batched as appropriate.

Each agent executes all web searches for its phase and returns ONLY a list of story items: title, URL, 1-line summary, and story_id. No analysis, no ranking - that happens in Phase 16.

Collect results from all research agents before proceeding to Phase 16. Do not skip phases - missing a phase means missing an entire digest section.

| Phase | Topic | Skip unless focus includes |
| --- | --- | --- |
| 2 | Technical research (models, papers, frameworks) | technical |
| 3 | Business research (funding, acquisitions, launches) | business |
| 4 | Engineering impact (dev tools, workflow, job market) | engineering |
| 5 | Leadership research (strategy, org transformation) | leadership |
| 6 | GitHub trending AI repos | technical |
| 7 | AI tools for professionals (9 domains) | all (always run) |
| 8 | AI application domains (7 verticals) | all (always run) |
| 9 | AI safety & ethics | all (always run) |
| 10 | Open source AI ecosystem | technical |
| 11 | AI infrastructure & hardware | technical |
| 12 | Regional AI developments | all (always run) |
| 13 | YouTube AI videos | all (always run) |
| 14 | Cool & thought-provoking research | all (always run) |
| 15 | Newsletter & blog aggregation | all (always run) |

### Phase 16: Synthesis (CRITICAL — Dedup BEFORE Digest)

All deduplication happens here before generating the digest.

**Step 1: Generate story IDs** for all collected items.

Normalized format — lowercase, hyphen-separated, company/product + action + key detail:

- "Falcon-H1R 7B release" → `falcon-h1r-7b-release`
- "xAI raises $20B" → `xai-20b-funding`
- "Simon Willison on sandboxes" → `simonwillison-sandboxes-post`

**Step 2: Deduplicate within session** — remove same event from different URLs.

**Step 3: Deduplicate against history** — use in-memory `covered_ids` and `covered_urls` from Phase 1. DO NOT re-read or update the file.

Filter out stories where:

1. Exact story_id match in `covered_ids`
2. Similar story_id (same product/company + same action)
3. Exact URL match in `covered_urls`
4. Same announcement, different angle

**Step 4: Rank** by source credibility (tier 1 > tier 2 > tier 3), engagement, relevance.

**Step 5: Categorize** into template sections and select Top 5 from filtered content.

**Step 6: Completeness check** — compare categorized items against the Length Guidelines table in [references/output-template.md](references/output-template.md).
If any section is below its target minimum, return to the corresponding research phase and run additional searches from [references/search-patterns.md](references/search-patterns.md) to fill the gap.
Every section in the template must have content before proceeding.

### Phase 17: Generate Digest

1. Load [references/output-template.md](references/output-template.md)
2. Fill sections with filtered items from Phase 16 only
3. Format: `- [ ] **[Title]** — [1-line summary] [Source: URL]`
   - Checkbox `- [ ]` on ALL story items with source URLs (renders as Notion task)
   - NO checkbox on: Top 5 summary, prose bullets, Action Items, Things to Explore, Connections, Sources section
4. If fewer than 5 stories after filtering, note "Light news day"

### Phase 18: Save Digest Files

DO NOT update `.covered-stories` in this phase — wait for verification.

**Step 1: Save to Notion**

Skip this step if `notion_page_id` was not resolved in Phase 1 (archive-only mode).

Load Notion tool via ToolSearch (`select:mcp__notion__notion-create-pages`), then create page:

- Parent page ID: use `notion_page_id` resolved in Phase 1
- Title: `🤖 AI Digest {YYYY-MM-DD}`
- Content: Full digest markdown (excluding H1 title)

If page creation fails, warn the user and continue — the archive copy in Step 2 still provides value.

**Step 2:** Write archive copy to `$DATA_DIR/ai-digest-{YYYY-MM-DD}.md`

**Step 3:** Update `$DATA_DIR/.last-run` with today's date (YYYY-MM-DD).

### Phase 19: Duplicate Verification

Spawn a `general-purpose` verification agent to check today's digest against:

1. `$DATA_DIR/.covered-stories` (should NOT include today's stories yet)
2. Last 3 digests from `$DATA_DIR/`

Agent checks for: exact duplicates, near duplicates (same company + similar action within 7 days), URL duplicates, topic fatigue (same topic 3+ times in past week).

**On duplicate detection:**

- REMOVE flagged: edit both Notion and archive copies, then proceed to Phase 20
- Borderline only: keep stories, add footer note `*Verification: {N} borderline items retained*`
- All clear: proceed to Phase 20

### Phase 20: Update Covered Stories (FINAL)

Only after Phase 19 passes. Append to `$DATA_DIR/.covered-stories` for each story in final digest:

```text
{date}|{story_id}|{url}
```

Keep file under 300 lines — trim oldest from top if over.

## Output Requirements

- Emojis for section headers (per template)
- Bullet points over paragraphs
- All items must have source URLs
- Top 5 stories section required
- Personal takeaways with actionable items
- Coverage period in header (e.g., "Coverage: Jan 25 - Jan 28 (3 days)")

## Error Handling

- If WebSearch fails for a source, log and continue with others
- Minimum viable digest: at least 5 items total
- If < 5 items, expand date range by 1 day and retry
- Only update `.last-run` on successful digest generation

## Newsletter Integration

Story items use `- [ ]` checkbox format for newsletter curation. User checks stories in Notion → `/ai-newsletter` extracts checked items.

## Example Invocations

<example>
Default digest with all focus areas:

```bash
/ai-digest
```
</example>

<example>
Focus on a single area:

```bash
/ai-digest --focus technical
/ai-digest --focus business
```
</example>

<example>
Explicit Notion page or archive-only mode:

```bash
/ai-digest --notion-page-id 12345678-abcd-1234-efgh-123456789abc
/ai-digest --no-notion
/ai-digest --focus technical --no-notion
```
</example>
登录后下载/点赞/收藏 ❤ 2 | ★ 0
ai-daily-digest

目录结构

SKILL.md

举报内容

提示
