SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsmorning-intel

morning-intel

Writing & research

View repository

Updated 10-16 06:52

About this skill

morning-intel

Many thin deterministic fetchers, one synthesis pass. The fetchers are plain Python (no AI in the loop; every failure degrades to a status flag instead of crashing the brief); Claude does Gmail triage, X gap-fill, and the editorial synthesis.

First run? If ~/.claude/skills/morning-intel/profile.md does not exist, the skill pack's installer hasn't run — see the repo README / CLAUDE.md.

Step 1 — Run the fetch layer (parallel)

Run all five in parallel (independent; failures don't block each other):

python ~/.claude/skills/morning-intel/scripts/fetch_smol.py
python ~/.claude/skills/morning-intel/scripts/fetch_hn.py
python ~/.claude/skills/morning-intel/scripts/fetch_youtube.py
python ~/.claude/skills/morning-intel/scripts/fetch_github.py
python ~/.claude/skills/morning-intel/scripts/fetch_changelogs.py

(Windows headless/cron wrapper: scripts/run_all.ps1 — same scripts via Start-Job, 6min timeout, never throws. On macOS/Linux just background the five commands with & + wait.)

Each writes <VAULT>/inbox/research/morning-intel/cache/YYYY-MM-DD/<source>.json with {status: ok|stale|error|mock, error, items}. Always exit 0. VAULT is the AGENTIC_OS_VAULT entry in ~/.claude/.env (defaults to ~/the-vault). github's "items" embeds the full github-trending report (it delegates to that skill's fetch.py, installed alongside this one).

Source What it covers Fallback when status != ok
smol X (544 accts) + Reddit + Discords, pre-aggregated daily lean harder on Step 3 X searches
hn front page + last-24h AI stories by points skip section, note in status table
youtube last-7d videos for 4 queries, views + views/subs outlier ratio fall back to yt-search skill (slow) or skip
github 7d/30d new repos + 24h/30d star velocity read yesterday's github-trending file
changelogs Codex changelog + Claude Code releases (raw text) skip, note it

Step 2 — Gmail triage (optional, read-only, ~5 tool calls)

Requires the Gmail connector (claude.ai connectors). If unavailable, skip and mark gmail: unavailable in the status table — the rest of the brief stands.

  1. search_threads with query newer_than:1d in:inbox, pageSize 50.
  2. Dedupe by THREAD; use the latest message date per thread (old threads with fresh replies match newer_than:1d, so classify off the newest message).
  3. Snippets classify ~80% of mail. Call get_thread ONLY for: transactional/ operational senders (billing, quotas, site notifications), and personalized inbound (a human who clearly knows the user's work).
  4. Categories: needs-reply / opportunity / newsletter-worth-reading / FYI / ignore. One why-care sentence each.
  5. Standing rule: quota/billing/site-breakage emails surface at the TOP of the brief as 🚨 operational items, above everything else.
  6. Read-only. Never label, draft, archive, or click tracking links.

Step 3 — X/Twitter gap-fill (WebSearch battery)

smol.ai covers most X discourse but lags ~1-2 days. Fill the gap with a fixed WebSearch battery (direct x.com fetches fail with HTTP 402; Nitter is dead — do not try them):

  • site:x.com AnthropicAI OR claudeai OR ClaudeDevs <topic keywords>
  • site:x.com bcherny OR _catwu OR alexalbert__ claude code
  • site:x.com sama OR gdb OR OpenAIDevs OR Codex_Changelog
  • Anthropic OR "Claude Code" announcement news <today's date>
  • OpenAI OR Codex announcement news <today's date>

Snippets return tweet text but NOT engagement counts — don't invent numbers. Date-check results: site:x.com recency ranking is unreliable, so verify a tweet is actually fresh (snowflake-ID magnitude or corroborating dated coverage) before presenting it as last-24h. Skip this step entirely if WebSearch is unavailable; smol + hn still carry the news sections.

Step 4 — Synthesize the brief

Read all cache JSONs + Gmail triage + X results + the user's ~/.claude/skills/morning-intel/profile.md (who they are, what formats/goals the So What section should target), then write:

<VAULT>/inbox/research/morning-intel/YYYY-MM-DD-intel.md

Idempotent: re-running the same day regenerates (overwrites) the file.

Section schema (keep headings stable):

# Morning Intel - YYYY-MM-DD (Weekday)

## TL;DR
3-6 bullets. The single most important story first. Any 🚨 operational email
items belong here too.

## Top Story
The one dominant narrative, 1-2 paragraphs, with links + the quotable stat.

## AI News (last 24h)
Grouped: Anthropic/Claude · OpenAI/Codex · Everyone else. Source links on
every claim. Changelog deltas fold in here.

## Hacker News Pulse
Front-page reads + last-24h AI stories worth knowing (points/comments).

## YouTube Radar
Table: title, channel (subs), views, ratio. Flag outliers (ratio >= 0.3) —
those are the angles resonating. Note formula convergence (multiple channels,
same hook). Non-English videos appear despite the US region hint — lead with
English results but call out international outliers as a lane signal.

## GitHub Radar
Top new repos (7d/30d) + fastest-growing, AI-relevant picks only — pull from
the embedded github-trending report. One line each on why it matters to the
user (per profile.md).

## Inbox (last 24h)          <- only if Gmail connected
🚨 Operational first, then 💰 opportunities/needs-reply, then FYI. Ignore-tier
gets one collapsed line with count, not a listing.

## So What - Action Plan
The payoff section, driven by profile.md. For content creators: 3-5 ranked
video/post ideas with working title, hook, which findings each rides, and
timing urgency (ride-now with expiry date vs evergreen). For builders: what to
try, adopt, or watch. Angles that extend something the user already does beat
cold ones.

## Source Status
| source | status | note |
One row per source incl. gmail + x-search. Honest about failures.

_Generated at HH:MM local_

Step 5 — Deliver

Chat summary: TL;DR + top story + the #1 action angle + any 🚨 inbox items. Link the vault file. Don't paste the whole brief into chat.

Cost + cadence notes

  • YouTube: ~410 quota units/run of the 10k/day free Data API quota — fine daily, don't hammer manually.
  • GitHub: unauthenticated works; set GITHUB_TOKEN in ~/.claude/.env for burst headroom.
  • Reddit has NO direct free path (all routes 403) — smol.ai is the Reddit proxy. Direct Reddit would need a free Reddit script-app OAuth token (future work).
  • Optional upgrade: SOCIALDATA_API_KEY (socialdata.tools, ~$0.20/1k tweets) would add real X engagement numbers for a watchlist; not wired yet.
  • Keep this on a LOCAL scheduler (Task Scheduler / cron / launchd) if the Gmail step is enabled — personal inbox data shouldn't route through cloud-scheduled agents.