daily-ai-news

name: daily-ai-news description: Generate a daily Thai-language AI news brief covering up to 5 stories from the last 24 hours (rolling window, Asia/Bangkok), with URLs deduplicated against yesterday's brief so the same story is never covered twice in a row, enrich it with three expert perspectives (professor / AI specialist / professional programmer), commit it to this repository via the GitHub connector, and push a TL;DR to LINE. Use this when the user asks for a "daily AI news brief", a "สรุปข่าว AI วันนี้", triggers this skill by name, or when it runs on schedule via Claude Web Routine.

Daily AI News Brief

End-to-end routine that produces one Markdown article per day at articles/YYYY-MM-DD-brief.md, commits it via the GitHub connector, then notifies LINE. Runs entirely inside Claude (Web / Routine) — no shell, no git CLI, no Bash.

Runtime contract

• Timezone: Asia/Bangkok. Compute YYYY-MM-DD from this TZ.
• Tools allowed: WebSearch, WebFetch, Read, Write, Edit, and the configured GitHub MCP connector.
• Tools FORBIDDEN: Bash, any shell, any git CLI. If you catch yourself reaching for Bash, stop — this skill must run on Claude Remote Routine where shell is unavailable.
• No fabrication: every news item MUST have a real, fetched URL from the trusted-source list in reference/trusted-sources.md. If you cannot verify a URL via WebFetch, drop the item.

Required environment

The Routine itself needs very little env now — LINE moved out to GitHub Actions (see Step 6). The skill reads only:

LINE_CHANNEL_ACCESS_TOKEN / LINE_TO are no longer read by the skill — they live in GitHub repo secrets and are consumed by .github/workflows/line-notify.yml. If you still see them in an older prompt, ignore them; the Routine doesn't need them.

Claude Web Routine's Cloud Environment feature does not reliably inject values into the model's prompt context (observed empirically in this deployment). That's fine — GitHub vars have a committed defaults fallback, and LINE secrets now live on a surface (GH secrets) that actually works.

Env resolution order (apply per variable, first hit wins)

1. Inline in the skill invocation prompt. The caller may have written GITHUB_OWNER = thannob directly in the run prompt. Use that.
2. reference/defaults.json — read with the Read tool. This is the canonical fallback for the three GITHUB_* vars.
3. Missing → hard abort.

Do not invent values. Do not treat literal placeholder strings like {{GITHUB_OWNER}} or $GITHUB_OWNER as real values — those are signs substitution failed and the caller needs to either paste real values or let the defaults take over.

Step 0 — Preflight (fail fast, log loudly)

1. GitHub connector check. Confirm the GitHub MCP connector is connected (look for tools whose names start with mcp__github / mcp__Github / similar — e.g. create_or_update_file, push_files, get_file_contents). If none are available:
   • Print: ABORT: GitHub connector is not connected. Enable the GitHub MCP connector in Claude settings and re-run.
   • Stop immediately. Do not research, do not write any files.
2. Resolve env. For each variable in the table above, walk the resolution order. Read reference/defaults.json exactly once and cache it.
3. Print a resolution table so failures are obvious:

   Env resolution:
     GITHUB_OWNER   = thannob       (source: defaults.json)
     GITHUB_REPO    = dailyainews   (source: defaults.json)
     GITHUB_BRANCH  = main          (source: defaults.json)
   

4. GitHub gate. If GITHUB_OWNER or GITHUB_REPO is ***missing*** → abort with a clear log.
5. Date. Compute TODAY = YYYY-MM-DD in Asia/Bangkok. Use this string for the filename, commit message, and article body.

Step 1 — Research (last 24h, deduplicated against yesterday)

Open reference/trusted-sources.md and treat it as the allow-list. Prefer sources from that list; do not invent outlets.

1a. Load yesterday's URLs (so we don't repeat ourselves)

Compute YESTERDAY = TODAY − 1 day in Asia/Bangkok.

Use the GitHub connector's get_file_contents to read articles/{YESTERDAY}-brief.md on GITHUB_BRANCH. Three outcomes:

• File found. Extract every URL from the body using a regex over inline-link syntax, i.e. all matches of \[[^\]]+\]\((https?://[^)]+)\) plus any plain URLs. Build a set YESTERDAYS_URLS. Lowercase host for the comparison; preserve the full URL otherwise (don't strip query strings or fragments — different paths are genuinely different stories).
• File not found (first-ever run, or yesterday's run was skipped). Set YESTERDAYS_URLS = {} (empty).
• Connector error other than 404. Surface the error and continue with YESTERDAYS_URLS = {}. Don't block the run on a stale-dedup-set worry.

Print the size: Yesterday's URL count: N.

1b. Gather candidates with WebSearch

Use WebSearch with rolling-24h queries:

• AI news (and Thai: ข่าว AI, ปัญญาประดิษฐ์). Add a date hint when supported (e.g. qdr:d on Google).
• Targeted per trusted source: site:techcrunch.com AI, site:blognone.com AI, etc.

Capture for each candidate: URL, title, publisher (inferred from domain vs trusted-sources list), and the search-result snippet verbatim — including any timestamp string the search engine surfaces (e.g. "2 hours ago", "April 27, 2026", "วันนี้").

1b. Verification — tiered (handles egress-blocked runtimes)

Each story is assigned a verification tier. Do a quick WebFetch probe on one trusted URL (e.g. https://example.com) at the start of this step to detect whether WebFetch works at all in this runtime. Label the runtime:

• WEBFETCH_OK — probe returned 2xx.
• WEBFETCH_BLOCKED — probe returned 403 / network error (typical in Claude Web Routine today).

Then for each candidate:

Never cite a URL that you could not at least see in a WebSearch result for a trusted-source domain. Search engines don't invent URLs, so a URL present in search results is real. The risk is staleness of the snippet, not fabrication — mitigate by never asserting more than the snippet says.

1b-bis. Two filters — last 24h AND not in yesterday's brief

Apply both filters before selection. Drop on either failure.

Filter A — Rolling 24h freshness. A candidate passes only if its publish timestamp falls within [now − 24h, now] in Asia/Bangkok.

Filter B — Not in yesterday's brief. If the candidate's URL is in YESTERDAYS_URLS (loaded in Step 1a), drop it — even if it's still trending. We're publishing today's reporting, not re-running yesterday's coverage.

This is URL-level dedup, not topic-level. Different reporting on the same evolving story (e.g. a follow-up on the Google–Anthropic deal at a different URL) is fair game — that's genuinely new content. Same URL = same article = drop.

Edge case: if a story's snippet body references an event from yesterday but the URL itself is fresh-today reporting on it, the URL passes Filter B (URL not in yesterday's set) and Filter A (24h timestamp). Include it. The dedup is precise: the specific article wasn't in yesterday's brief, even if the topic was.

1b-tris. Hard rule — DO NOT bend the 24h filter

Filter A is a hard gate, not a guideline. Forbidden patterns observed in past runs that this rule explicitly prohibits:

• ❌ "I couldn't find 5 stories from today, so I included items from the past several days." → NO. Ship fewer; never older.
• ❌ "Most-recent indexed stories on trusted-sources.md happen to be from 3-5 days ago, so I'll use them." → NO. "Most recent on the index" ≠ "within 24h." If indexers haven't crawled today's news yet, that's a search-recency issue — try other queries, other domains, broader Thai sources. If still empty, ship the empty stub.
• ❌ "The 24-hour window could not be guaranteed via search snippets alone, so older items were included." → NO. That note is the sound of the contract breaking. If you can't guarantee freshness, you can't include the item. Period.
• ❌ Including an item with Published: per search snippet (อ้างอิงข้อมูล [last quarter / last month]). The dated content is what's being aggregated, not what's being reported today. Drop.

Before writing sources.md, re-check every selected story's timestamp against now − 24h Asia/Bangkok one more time. Any item that fails this check at the second pass is dropped silently — no apology note, no inclusion-with-caveat. If this leaves you with 0 stories, write the stub and commit.

The empty-day signal is correct output. A padded brief is broken output that erodes trust in every future run.

1c. Select up to 5

Select up to 5 stories that passed both filters (24h fresh AND not in YESTERDAYS_URLS). Mix: aim for ≥1 Thai-language source and ≥3 international sources when supply allows. Prefer Tier 1 over Tier 2 when both are available for the same candidate. Prefer primary announcements over commentary.

If fewer than 5 candidates pass, ship what you have — even 1, even 0. Do not loosen either filter to fill the count.

If 0 stories pass → write a one-line stub article (per Step 5a NO-OP rules) explaining whether the shortfall came from (a) the 24h freshness filter, (b) the dedup filter, or (c) both, and commit. The empty-day signal is itself information.

1d. Write reference/sources.md

Overwrite the file with this template — Verification, FreshnessCheck, and DedupCheck lines are all required:

# Sources — {TODAY}

Generated: {TODAY} (Asia/Bangkok)
Runtime: {WEBFETCH_OK | WEBFETCH_BLOCKED}
Freshness window: rolling 24h (Asia/Bangkok)
Dedup against: articles/{YESTERDAY}-brief.md ({N} URLs loaded)

1. **{Headline}**
   - Publisher: {Publisher}
   - URL: {final URL}
   - Published: {explicit timestamp or relative phrase as it appeared}
   - FreshnessCheck: ✅ within last 24h via {evidence — e.g. "snippet '3 hours ago'", "URL slug 2026-04-27", "publishedTime 2026-04-27T08:14"}
   - DedupCheck: ✅ URL not in YESTERDAYS_URLS ({reason — e.g. "domain matches yesterday but path differs", "host not in yesterday at all")
   - Verification: {Tier 1 — WebFetch | Tier 2 — WebSearch snippet}
   - Summary: {1–2 sentences, strictly from fetched body or snippet}

2. ...

If you saw promising candidates but they failed either filter, list them in a "Dropped" block for traceability:

## Dropped
- {URL} — Filter A (>24h): "URL slug /2026/04/24/... is 3 days old"
- {URL} — Filter B (dedup): "appeared in articles/{YESTERDAY}-brief.md"

If you found <5 items, end with:

> Note: {N} items passed both filters this run. Of {M} candidates, {X} failed Filter A, {Y} failed Filter B.

If 0 stories passed both filters → skip Step 2–4, write a minimal article (Step 5 still commits) explaining whether the shortfall came from (a) Filter A (no fresh news), (b) Filter B (everything available was already covered yesterday), or (c) both. The empty-day signal is itself information.

Step 2 — Draft the article

Create the first draft in memory (don't commit yet). Structure:

# สรุปข่าว AI ประจำวันที่ {TODAY}

> TL;DR
> - {bullet 1 — one sentence}
> - {bullet 2}
> - {bullet 3}

## ข่าวเด่น 24 ชั่วโมงที่ผ่านมา

### 1. {Headline}
{2–4 sentences. Link inline: [{Publisher}]({URL}). Every factual claim must trace to that URL.}

### 2. ...

Rules:

• Every story references its source URL at least once via inline markdown link.
• Do not invent quotes, numbers, dates, or names. If the source didn't say it (Tier 1 body) or if the snippet didn't say it (Tier 2), don't write it.
• Thai-first prose; technical terms can stay in English.
• Do NOT mention verification mode anywhere in the article body. No top banner, no italic tags, no footer line about Tier 2 / WEBFETCH_BLOCKED / search-snippet. The reader doesn't need any of that — they want news, not metadata about how it was sourced. Verification mode lives only in two places, both invisible to the reader: the commit message [verify=search|webfetch] tag, and reference/sources.md's Runtime: + per-story Verification: lines. That's plenty for an operator to audit; nothing belongs in articles/{TODAY}-brief.md itself.

Step 3 — Three perspectives

For each of the 5 stories, produce a short reaction from three personas. Write them to reference/perspectives.md (overwrite) using this layout:

# Perspectives — {TODAY}

## 1. {Headline}

**อาจารย์ (มหาวิทยาลัย):** {1–2 sentences — pedagogical framing, what students/readers should understand}
**ผู้เชี่ยวชาญด้าน AI:** {1–2 sentences — technical substance, caveats, what's genuinely new}
**โปรแกรมเมอร์มืออาชีพ:** {1–2 sentences — practical impact on day-to-day engineering work, tooling, cost}

## 2. ...

Keep each persona's voice distinct. No filler. No platitudes.

Step 4 — Rewrite (integrated)

Produce the final article body by weaving the three perspectives into each story rather than listing them as separate blocks. Append an Action items section at the end.

Target structure for the final file articles/{TODAY}-brief.md:

# สรุปข่าว AI ประจำวันที่ {TODAY}

> TL;DR
> - {bullet 1}
> - {bullet 2}
> - {bullet 3}

## ข่าวเด่น 24 ชั่วโมงที่ผ่านมา

### 1. {Headline} — [{Publisher}]({URL})
{2–4 sentences of reporting, then a paragraph that integrates the professor / AI-expert / programmer angles naturally. No "persona:" labels in the body.}

### 2. ...

## Action items

- **สำหรับอาจารย์/นักเรียน:** {1 concrete action}
- **สำหรับผู้เชี่ยวชาญ AI:** {1 concrete action}
- **สำหรับโปรแกรมเมอร์:** {1 concrete action}

---
_Generated by the `daily-ai-news` skill on {TODAY} (Asia/Bangkok)._

Save the final markdown as a string in memory — call it ARTICLE_BODY — for the commit step.

Step 5 — Commit via GitHub connector

Never use git CLI. Use the connector tool. The exact tool name depends on the connector build; common shapes:

• create_or_update_file with params: owner, repo, path, branch, message, content (plain UTF-8 string; connector handles base64 if needed), optional sha for updates.
• push_files with a files array.

5a. Idempotency check

Before writing, call the connector's get_file_contents (or equivalent) for articles/{TODAY}-brief.md on branch. Three outcomes:

• Not found. Create it. Proceed.
• Found, content byte-for-byte identical to ARTICLE_BODY. Skip the commit. Log: Run status: NO-OP (idempotent) — today's brief already committed at <existing SHA>. Capture that SHA as COMMIT_SHA and still proceed to Step 6 (LINE may have been skipped last time).
• Found, content differs. Update with the returned sha — this is a meaningful re-run (e.g. the earlier run was WEBFETCH_BLOCKED and this one has fresh data).

5b. Commit

1. Set:
   • path = articles/{TODAY}-brief.md
   • branch = GITHUB_BRANCH or "main"
   • message = "brief: {TOPIC} {TODAY} [verify={webfetch|search}]" where {TOPIC} is a ≤40-char summary of the day's dominant theme (e.g. OpenAI ships new agent API), and verify= reflects the runtime label from Step 1b.
   • content = ARTICLE_BODY
2. Call the connector's create-or-update tool.
3. If the response includes a commit SHA (typical field: commit.sha or sha), capture it as COMMIT_SHA. Otherwise attempt to read it back by calling the connector's get_file_contents / commit-list tool for the branch.
4. Form PERMALINK = https://github.com/{GITHUB_OWNER}/{GITHUB_REPO}/blob/{COMMIT_SHA}/articles/{TODAY}-brief.md. Pin to the SHA — do not use /blob/main/... (that link drifts).
5. If the commit call fails: print the full connector error, do not retry silently, do not proceed to LINE.

Step 6 — LINE push (handled externally)

This skill does NOT send to LINE. The WebFetch tool in Claude Web Routine accepts only (url, prompt) — no method, no headers, no body — so it cannot POST to api.line.me/v2/bot/message/push with a Bearer token. That's a tool-schema limit, not an egress or env issue.

LINE dispatch is instead handled by a GitHub Actions workflow at .github/workflows/line-notify.yml that triggers on every push to main affecting articles/*-brief.md. The workflow:

• Reads repo secrets LINE_CHANNEL_ACCESS_TOKEN and LINE_TO.
• Extracts the title + TL;DR bullets from the newly committed brief.
• Builds a permalink pinned to the commit SHA.
• POSTs to api.line.me with curl (full HTTP capability).
• Fails loudly in the Actions UI on non-200.

What the skill does here:

1. Do nothing that talks to LINE directly — no WebFetch, no markers, no retries.
2. Print one line for the operator: LINE: dispatched by .github/workflows/line-notify.yml (see Actions tab for delivery status).
3. If the commit in Step 5 was a NO-OP (idempotent), the workflow will NOT fire — because no push happened. This is correct: no new content means no new notification.

If the maintainer wants to force a re-notify for an already-committed brief, they can run the workflow manually via gh workflow run line-notify.yml -f file=articles/{TODAY}-brief.md or the GitHub UI.

Step 7 — Final report

Print a compact summary to the user / routine log:

✅ Committed: articles/{TODAY}-brief.md @ {COMMIT_SHA_short}
   {PERMALINK}
LINE: dispatched by .github/workflows/line-notify.yml (see Actions tab for delivery status)

If Step 5a decided NO-OP (identical content already on main), say so instead:

Run status: NO-OP (idempotent) — today's brief already at {COMMIT_SHA_short}
   {PERMALINK}
LINE: not re-notified (no new commit → workflow does not fire)

Error-handling summary

Files this skill touches

• reference/sources.md — overwritten each run (working artifact)
• reference/perspectives.md — overwritten each run (working artifact)
• articles/{TODAY}-brief.md — the single committed output

reference/trusted-sources.md and reference/defaults.json are read-only for this skill.