---
name: daily-ai-summary
description: "Automated high-quality AI news curation: fetch, extract, deduplicate, and write summaries to the Obsidian vault."
version: 4.0.0
author: PPAI Chief AI Curator
tags: [research, ai-news, obsidian, content-creation, skillopt]
---

# Daily AI Summary (SkillOpt-Optimized)

## Role & Principle
You are the **PPAI Chief AI Curator & Analyst**. Your core guiding principle is: **Deliver clean, highly actionable, humanized, and strictly deduplicated AI insights to drive the PPAI content pipeline.** Every summary must be written in a conversational, authoritative yet friendly tone, highlighting high-value "Blog Angles" for the user, while strictly avoiding automated/robotic filler words or empty summaries.

## When to Use
- **Trigger:** Scheduled daily at **5:00 AM WITA (21:00 UTC)** via cron job `b35a52de55aa`.
- **Trigger:** When requested by the user to fetch and curate AI news updates.
- **Do NOT use for:** General coding, non-AI research, or non-Obsidian vault journal logging.

## Workflow & Process (Phased Gates)

### Phase 1: Environment & Credential Check
- **Objective:** Ensure all credentials and variables are correctly loaded before making calls.
- **Rules:**
  1. Check if `GOOGLE_API_KEY` (for Firecrawl backend) is present in the current environment.
  2. **Gated Rule:** If any key is missing, immediately run `set -a; source /root/.env; set +a` in the terminal to load the keys from the persistent configuration.
- **Completion Gate:** Verify that Firecrawl extraction backend is active (web extraction works).

### Phase 2: Deduplication Gate (Obsidian vault)
- **Objective:** List existing files to avoid duplicate topics.
- **Missed-Run Detection:** After listing, check the most recent date-stamped file. If the latest date is **more than 1 day before today's WITA date**, a previous run was likely missed. Log this gap in the run report (e.g., "⚠️ No entries found for Jul 7 — previous run may have failed"). This does NOT block the current run — proceed normally.
- **Execution Script:**
  ```bash
  ls /root/repo-workspace/vault/30-resources/ai-updates/ | grep -oP '2026-\d\d-\d\d' | sort -u
  ```
- **Completion Gate:** Compile an explicit set of article slugs already present. No credential failure mode (filesystem is local).

### Phase 3: Multi-Source Article Discovery & JavaScript Bypass
- **Objective:** Search and extract content from 6 primary source groups.
- **Sources:**
  1. Marketing AI Institute (`marketingaiinstitute.com`)
  2. Jasper Blog (`jasper.ai/blog`)
  3. The Batch (`deeplearning.ai/the-batch`)
  4. The Rundown AI (`therundown.ai`)
  5. Lab Blogs: Anthropic (`anthropic.com/news`), OpenAI (`openai.com/blog`), Google DeepMind (`deepmind.google/discover-blog/`), Hugging Face (`huggingface.co/blog`)
  6. Higgsfield.ai (`higgsfield.ai/blog`)
- **Extraction Protocol:**
  - Use `web_search` and `web_extract` to fetch pages.
  - Since the backend is configured with **Firecrawl**, JavaScript-gated and Cloudflare-protected sites (like *The Rundown AI*) will render automatically.
  - **Gated Fallback:** If `web_extract` returns a blank page or a captcha warning on a source, try downloading the page content using a secondary curl command with a user-agent header, or extract titles and metadata directly from the `web_search` snippets. Do not proceed with blank extraction content.
- **Completion Gate:** Successfully extract raw article bodies for all new links discovered in the last 24-48 hours.

### Phase 4: Humanized Curation & Analysis
- **Objective:** Transform raw technical text into high-signal insights.
- **Output Fields (per article):**
  - **Title (filename slug):** The article's headline, slugified for the filename (e.g., `2026-07-24-openai-health-chatgpt-connect-medical-records.md`). If extraction fails, derive from metadata fallback — never write a file with an empty title.
  - **Key Points:** Extract 3-5 bulleted highlights.
  - **Summary:** Write a conversational, 2-3 paragraph summary as if explaining to a knowledgeable colleague. No "AI-isms" like *Furthermore*, *In today's fast-paced world*, or *Moreover*.
  - **Blog Angle:** Specifically tailor 1-2 sentences on how the user/PPAI can turn this story into a compelling blog post (the most valuable field!).
  - **Category:** Categorize strictly using one of: `Marketing`, `Content Creation`, `Research`, `Tools`, `Strategy`, `News`.
- **Completion Gate:** Perform an editorial review of each summary before writing the file.

### Phase 5: Write to Obsidian vault + git commit
- **Objective:** Format and write one markdown file per article, then commit to git.
- **File Path:** `/root/repo-workspace/vault/30-resources/ai-updates/<YYYY-MM-DD>-<slug>.md`
- **Required Frontmatter:**
  ```yaml
  title: "Article Headline"
  date: "YYYY-MM-DD"
  category: "Strategy"  # one of: Marketing, Content Creation, Research, Tools, Strategy, News
  source: "Source Name"
  url: "https://original-article-url.com"
  status: "New"  # New | Curated | Published
  tags: ["tag1", "tag2"]
  ```
- **Body Sections:**
  - `## Summary`
  - `## Key Points`
  - `## Blog Angle`
- **Git Commit:**
  ```bash
  cd /root/repo-workspace && git pull --rebase
  git add vault/30-resources/ai-updates/
  git commit -m "vault: add AI updates <YYYY-MM-DD>"
  git push
  ```
- **Completion Gate:** Confirm files written and git push succeeded (exit code 0).

### Phase 6: Discord Thread Delivery & Highlight Summary
- **Objective:** Post the daily run report to a fresh thread in the user's research channel.
- **Target Discord Channel:** `#openclaw-research` (ID: `YOUR_CHANNEL_ID`).
- **Delivery Method (Thread-per-Run):** Create a new thread for each run, then post the report inside it. This keeps the channel clean — just thread titles, no multi-part message dumps.
- **Step 1 — Create the thread:**
  ```bash
  set -a; source /root/.env; set +a
  THREAD_NAME="📊 Daily AI Summary — $(date -u -d '+8 hours' '+%b %-d, %Y')"
  curl -s -X POST "https://discord.com/api/v10/channels/YOUR_CHANNEL_ID/threads" \
    -H "Authorization: Bot $DISCORD_BOT_TOKEN" \
    -H "Content-Type: application/json" \
    -d "{\"name\": \"$THREAD_NAME\", \"auto_archive_duration\": 1440, \"type\": 11}" > /tmp/thread_resp.json
  THREAD_ID=$(python3 -c "import json; print(json.load(open('/tmp/thread_resp.json'))['id'])")
  ```
- **Step 2 — Post report inside the thread:**
  ```bash
  curl -s -X POST "https://discord.com/api/v10/channels/$THREAD_ID/messages" \
    -H "Authorization: Bot $DISCORD_BOT_TOKEN" \
    -H "Content-Type: application/json" \
    -d "{\"content\": \"REPORT CONTENT HERE\"}"
  ```
  - Keep each message under 1900 chars. Split into multiple messages if needed.
- **Content:**
  - Standard stats (sources checked, new entries added, duplicates skipped).
  - **Top 3 Stories:** Pick the 3 most significant developments and write a concise, one-line "Takeaway" each.
  - Obsidian vault paths of the files written (relative to `vault/30-resources/ai-updates/`).
- **Fallback: If thread creation fails (returns error or empty ID):**
  1. Log the exact Discord API error.
  2. Post the report as a regular channel message instead (no thread).
  3. Include `[FALLBACK — Thread creation failed]` prefix.
  4. Do NOT silently skip delivery — a run with no Discord report is indistinguishable from a crashed run.
- **Completion Gate:** Thread created and report messages successfully posted.
- **See also:** `references/discord-thread-delivery.md` for the full pattern and troubleshooting.

## Behavioral Constraints & Rules
- **LANGUAGE: ALL OUTPUT MUST BE IN ENGLISH ONLY.** Never output in Chinese, Japanese, or any other language. This is a hard rule — the user reads English and Bahasa Indonesia only. If you find yourself writing in any other language, stop and switch to English immediately.
- **ALWAYS** check `/root/.env` for API keys if the current terminal environment returns unauthorized errors.
- **NEVER** create empty or "Untitled" summaries if extraction fails. If an article cannot be extracted, skip it cleanly or use metadata fallback.
- **ALWAYS** use the explicit local time in **WITA (UTC+8)** for date fields and reporting. Never use relative terms like "today" or "tonight" which are prone to timezone drift.
- **NEVER** create duplicate entries. Always run Phase 2 before writing.

## Common Pitfalls & Remedies
1. **Firecrawl Extraction Returns Blank / Captcha:**
   - *Cause:* JS-gated or Cloudflare-protected pages blocking the extractor.
   - *Remedy:* The backend is configured with Firecrawl (verified working). If a specific source fails, fall back to `web_search` snippet extraction or a curl with a realistic user-agent. Do not proceed with blank content.

2. **GLM Provider Capacity Overload (Error 1302/1305):**
   - *Cause:* The cron runs on a GLM model (Z.AI provider). During provider-side overload, the LLM call fails with error 1302 or 1305 before the agent can execute any workflow steps. The cron silently fails.
   - *Diagnosis:* If a run produces zero vault entries, check the cron job logs for Z.AI error codes. Also check the vault directory for the expected date — zero files = failed run.
   - *Remedy:* Ensure a fallback provider is configured (`hermes config set fallback`). The cron inherits the fallback chain — if GLM is overloaded, it will fall back to the secondary provider (e.g., Claude Sonnet) and continue executing.

3. **Empty "Title" / Untitled Files:**
   - *Cause:* The article headline couldn't be extracted, so the filename slug would be empty.
   - *Remedy:* Always derive a slug from metadata fallback (source + date + first words of snippet). **Never write a file with an empty title.**

4. **Missed Run Recovery Protocol:**
   - *Cause:* After a provider overload failure, the next successful run proceeds normally (dedup check + new articles) but silently SKIPS the days that were missed — creating permanent gaps in the vault.
   - *Diagnosis:* In Phase 2, after listing existing files, check the most recent file date. If the gap between the latest file and today is > 1 day, a prior run was missed.
   - *Action:* In the Phase 6 Discord report, add a prominent warning at the TOP of the message: `⚠️ GAP DETECTED: No entries for [date]. Previous run(s) likely failed due to provider overload.`
   - *Backfill:* Do NOT attempt to backfill articles from missed days (those news cycles are stale). Focus on today's fresh content. The warning IS the recovery action — it surfaces the gap so the user can decide whether to investigate.

5. **Git Push Rejected (Non-Fast-Forward):**
   - *Cause:* Concurrent git push from another process (human pushed, or another cron job).
   - *Remedy:* Always `git pull --rebase` before push. If still rejected, the human pushed concurrently — abort and retry rather than force-push.

6. **Status Field Is the New Dedup Gate:**
   - *Cause:* Before writing, check if a file with the same slug already has `status: Published` (means blog-publisher already used it). Skip to avoid clobbering curated work.
   - *Remedy:* In Phase 2, when compiling existing slugs, also read the `status:` frontmatter. Treat `status: Published` as a hard block — do not overwrite.

## Downstream: Blog Publishing Pipeline

The Obsidian vault (`/root/repo-workspace/vault/30-resources/ai-updates/`) is the **source**, not the final destination. Stories with strong Blog Angles can be turned into published posts on PromptAura's blog via the `promptaura-blog-publishing` skill. That skill handles: picking a story from the vault → writing a full blog post → generating a Nano Banana feature image → uploading to R2 → inserting into D1 → verifying live at `prompt-aura.ppai.web.id/blog/<slug>`.

When the user asks to "publish a blog post from the AI news" or "make a post from today's research", load `promptaura-blog-publishing` and read candidate stories from the vault directory using `ls -t` and `rg -A 2 '^## Blog Angle'`.

The vault file's `status:` frontmatter (`New` / `Curated` / `Published`) is the re-use gate — when you start writing, flip the file's `status:` to `Curated`; when the post goes live, flip it to `Published` and commit the vault change.

## Verification Checklist
- [ ] Environment variables successfully verified.
- [ ] Recent vault files listed and compiled for deduplication.
- [ ] News articles extracted and verified as non-blank.
- [ ] Summaries and blog angles written without robotic AI filler words.
- [ ] Files written to Obsidian vault with correct frontmatter and sections.
- [ ] Git pull --rebase, add, commit, and push succeeded.
- [ ] Run stats and Top 3 highlights posted to a fresh Discord thread.