/ads update <platform|all> — Self-Refreshing Platform References

This skill regenerates per-platform reference files with what has actually changed in the last 30 days across each ad platform. Sources include Reddit, Hacker News, official platform changelogs, and the open web.

Powered by an adapted version of last30days-skill (MIT, by Matt Van Horn). See scripts/lib/THIRD_PARTY_NOTICES.md for full attribution.

Step 0 — MANDATORY: Cost confirmation gate

Before doing anything else, you MUST:

1. Read the contents of ads/references/update-cost-warning.md and print it verbatim to the user.
2. Compute the estimated token cost for the requested platform(s):
   • Single platform: ~50,000–150,000 tokens at default depth
   • all: ~500,000+ tokens at default depth
3. Call AskUserQuestion with this exact question and options:
   • Question: "Run /ads update <platform>? Estimated cost: ~<N>k tokens. Proceed?" (substitute <platform> and <N> with actuals)
   • Options:
     • Proceed (default depth) — run the requested mode at full cost
     • Switch to quick depth — same coverage, ~40% of the cost
     • Cancel — abort with no fetches
4. If the user picks Cancel or any answer other than Proceed / Switch to quick depth, abort immediately with a one-line message: "Cancelled. No sources fetched, no files written."

Do not skip this step. It exists to protect users on low-credit plans from accidentally burning their balance.

Step 1 — Argument parsing

Accept exactly one argument from this enum:

meta | google | tiktok | all

If the user invokes /ads update with no argument, ask them which platform via AskUserQuestion (don't default to all). If they pass an unknown value, list valid options and exit.

For all, loop the three platforms sequentially in this order: meta, google, tiktok. Sequential matters — parallel hits rate limits on the public APIs and makes cost unpredictable.

Step 2 — Load source set for the platform

Run this Bash command to fetch the source configuration:

python3 scripts/ads_sources.py --list <platform>

Parse the JSON output. You'll get:

• display_name — human-readable platform name for the digest header
• subreddits — list of Reddit subs to query
• changelog_urls — official vendor release-note pages (use WebFetch)
• search_queries — phrased to surface industry-press coverage (use WebSearch)
• hn_keywords — terms to query against Hacker News
• x_handles — optional, skip if no X/Twitter API key

Step 3 — Fetch sources

For each enabled source, fetch the last-30-days items. Depth-mode caps how many items per source:

Reddit (free, no key)

For each subreddit, run:

curl -s "https://www.reddit.com/r/<sub>/new.json?limit=25" -H "User-Agent: claude-ads/2.0"

Parse the JSON, keep only posts with created_utc within the last 30 days. Compute published_at from created_utc (Unix timestamp → YYYY-MM-DD). Capture title, permalink, score, num_comments, upvote_ratio, selftext (first 500 chars).

Hacker News (free, no key)

For each keyword in hn_keywords, run:

curl -s "https://hn.algolia.com/api/v1/search_by_date?query=<keyword>&numericFilters=created_at_i>$(date -u -v-30d +%s)&hitsPerPage=10"

Capture title, url, points, num_comments, created_at.

Official changelogs

For each URL in changelog_urls, use the WebFetch tool with this prompt: "Extract any release notes, changelog entries, or feature announcements from the last 30 days. Return as a list of items with: title, date (YYYY-MM-DD), one-line summary, URL anchor if available."

Web search (industry press)

For each query in search_queries (capped per depth setting), use the WebSearch tool. Filter results client-side: prefer hits from searchengineland.com, searchenginejournal.com, marketingland.com, martech.org, adweek.com, vendor blogs.

Step 4 — Score and dedupe (optional but recommended)

If you fetched 30+ items, run them through the vendored scoring lib for ranking. Otherwise, skip and order by published_at descending.

python3 -c "
import sys, json
sys.path.insert(0, 'scripts')
from lib import signals, dedupe
from lib.story import Story

raw = json.load(sys.stdin)
items = [Story(**r) for r in raw]
items = signals.annotate_stream(items)
items = dedupe.dedupe_items(items, threshold=0.7)
print(json.dumps([{'source': i.source, 'title': i.title, 'url': i.url, 'published_at': i.published_at, 'rank_score': i.rank_score} for i in items], indent=2))
"

Step 5 — Render the digest

Write the output to ads/references/<platform>-changelog-30d.md (overwrite if it exists). Use this exact structure:

# <Display Name> — Last 30 Days

_Generated <YYYY-MM-DD> by `/ads update <platform>` (depth: <quick|default|deep>)_

> **What this is:** the last 30 days of changes to <platform> ads — features, deprecations, policy updates — pulled from official changelogs, practitioner discussion, and industry press.

## Official Changelog Entries

- [<title>](<url>) — <YYYY-MM-DD> — <1-line summary>
- ...

## Industry Press

- [<title>](<url>) — <YYYY-MM-DD> — <publisher> — <1-line summary>
- ...

## Community Discussion (Reddit / Hacker News)

- [<title>](<url>) — <YYYY-MM-DD> — r/<sub> or HN — <1-line summary>
- ...

---

_Sources fetched: <N> Reddit, <M> HN, <P> official changelog URLs, <Q> web searches. Deduped via hybrid Jaccard at threshold 0.7._

If a section has no items, print: _No items found in the last 30 days from this source class._

Step 6 — Append to existing platform audit

After writing the changelog file, append (do not overwrite) a "Recent Updates" appendix to ads/references/<platform>-audit.md. Keep it short — top 5 most-impactful items only:

---

## Recent Updates (Last 30 Days)

_Auto-appended by `/ads update <platform>` on <YYYY-MM-DD>. See `ads/references/<platform>-changelog-30d.md` for the full digest._

1. <impact> — [<title>](<url>) — <YYYY-MM-DD>
2. ...

If the audit file already has a "Recent Updates" appendix from a prior run, replace it rather than stacking duplicates. Match by the marker line ## Recent Updates (Last 30 Days).

Step 7 — Confirm to the user

Print a one-line summary:

✓ /ads update <platform> complete — wrote ads/references/<platform>-changelog-30d.md (<N> items) and appended top 5 to ads/references/<platform>-audit.md.

For all mode, print this once per platform as you complete it, and a final summary line:

✓ /ads update all complete — refreshed 7 platforms.

Failure modes

• Reddit blocked / 429: skip that subreddit, note in output. Don't retry — public Reddit JSON is best-effort.
• WebFetch URL 404: skip that changelog URL, note in output.
• WebSearch returns nothing: skip that query.
• All sources empty for a platform: still write the file with "No items found in the last 30 days" sections — gives the user explicit confirmation that the run completed.
• User cancels mid-all: stop the loop. Don't roll back already-written files (those are still valid, just partial coverage).

Never silently fail. Always print which sources succeeded and which were skipped.