New skills, practical workflows, and agent-builder notes.

Use a curated open font library for Chinese covers and video titles. Separate image generation from typesetting, choose the right file format, and keep licenses with your workflow.

A practical guide to creator workflows: what each Skill produces, which tools and accounts it needs, and what to check before publishing.

Keep the selected Drive entry, its target and the exported artifact distinct. Avoid filename-based substitutions and choose download or export from the actual target's current metadata.

A locator that matches one element can still identify the wrong record. Scope browser actions to business identity, preserve ambiguity and verify the intended object after navigation.

Separate branch contribution from differences between current snapshots. Give a review agent pinned commit IDs, a comparison purpose and an explicit boundary for uncommitted work.

Keep navigation, local actions and form submission distinct even when controls share a visual style. Give a frontend agent an effect map and a keyboard acceptance checklist.

Treat archive extraction as a filesystem write, not a harmless preview. Compare selective inspection with bounded extraction, including path, resource and partial-failure policies.

Prevent unexplained total differences by defining numeric representation, rounding mode and rounding stage. Includes a small executed decimal fixture, not a financial-system benchmark.

Review how a PowerPoint slide is understood without its visual arrangement. Use a semantic outline, object mapping and a recipient-side check before accepting an AI-generated deck.

Design conditional web collection around retained response bodies, request variants and extractor versions. A 304 response should not become an empty record or a false change alert.

Separate Gmail thread IDs, message IDs and MIME reply headers. Build a draft-first reply workflow that preserves conversation context without silently expanding recipients.

Separate untrusted pull request execution from privileged repository actions. Review triggers, checkout sources and token scope before accepting an agent's CI repair.

Separate sitemap discovery, live-page checks and Google's indexed evidence. Build a URL cohort report that shows what is known without promising rankings or instant indexing.

Make the intended video, narration and subtitles explicit before export. Use a stream inventory and output checks instead of assuming an encoder will infer the editorial brief.

Review NOT IN and NOT EXISTS with explicit null policies. Keep a missing exclusion key from silently changing a reporting population or turning unknown identity into approval.

Review motion by its purpose and trigger. Keep loading, success and navigation feedback understandable when decorative animation is removed from an agent-built interface.

Choose Git restore, reset or revert by the state you intend to change. Preserve unrelated edits and review the exact rollback scope before an agent repairs a broken change.

Choose mock cleanup by the state that should survive. Keep call history, stub behavior and spied object methods separate before an agent adds a suite-wide fix.

Separate duplicate bibliography records from distinct versions and duplicate files. Give a research agent a reviewable merge plan before it cleans up your Zotero library.

Separate actionability from application readiness before extracting dynamic tables. Define the requested state, completion evidence and timeout outcome instead of adding arbitrary sleeps.

Separate property totals from visible query rows before explaining a traffic gap. Give SEO agents a scope record, coverage caveat and honest denominator for branded-query analysis.

Separate source clock time, UTC instants and business dates before grouping events. Keep DST uncertainty visible instead of letting a reporting agent silently move transactions.

Build a separate release copy of an agent-generated deck. Check hidden slides, notes and embedded content before treating a polished preview as approval to distribute the file.

Give video agents an audio delivery target, selected stream and final-file check. Separate loudness normalization from a mix that actually makes the narration understandable.

Separate a completed page walk from complete discovery. Give GitHub collection agents a coverage receipt, resumable checkpoints and an honest stopping rule.

The same image can need different text in a gallery, a button or a report. Review context, accessible names and chart explanations before bulk-generating descriptions.

A valid CSV can still become active spreadsheet content. Separate machine exports from human-facing workbooks and check cell types before handing an agent's report to a colleague.

Plan font availability, editing needs and a recipient-side check before accepting an AI-generated deck. An approved preview does not prove the shared PPTX will render the same.

Separate source timestamps, composition frames and delivery frame rate. Preserve cue timing when an agent changes export settings instead of treating fps as a cosmetic label.

Choose chunked pandas or DuckDB by the operation and result size, not CSV size alone. Keep a successful large-data query from becoming an oversized in-memory handoff.