Frontier Tracker

Use this skill to run a personal literature surveillance workflow. Keep every output concise, dated, and easy to act on.

Load the minimum context

• Windows encoding: Always run Python scripts with python -X utf8 on Windows to avoid GBK encoding errors.
• Read references/source-and-selection-rules.md when building or refreshing the journal pool.
• Read references/top-journal-families.md when the user asks for top journals, Nature/Science/Cell/PNAS family tracking, or when no ranking export is available.
• Run python -X utf8 scripts/build_top_family_watchlist.py --format json to turn the top-family seed list into a machine-readable watchlist; add --resolve-openalex only when ISSN/OpenAlex source IDs are needed.
• Run python -X utf8 scripts/scan_recent_papers.py --days 7 --update-state --output outputs/data/frontier_scan_<date>.json for repeatable top-family scans with DOI/OpenAlex de-duplication.
• After the scan, run python -X utf8 scripts/enrich_abstracts.py --input outputs/data/frontier_scan_<date>.json --output outputs/data/frontier_enriched_<date>.json --field new to fill missing abstracts via three-way fallback (OpenAlex → Crossref → Semantic Scholar).
• Then run python -X utf8 scripts/screen_by_profile.py --input outputs/data/frontier_enriched_<date>.json --output outputs/data/frontier_screened_<date>.json --profile references/<profile>.md --field both to classify papers into core / proxy / eco / noise tiers using profile keywords.
• Read references/example-profile.md when the user needs help creating a research profile.
• Read references/research-profile-template.md only when the user wants to replace or extend their profile.
• Read references/output-templates.md when producing a daily digest, weekly report, reading list, or reminder list.
• Read references/display-output-options.md when the user asks how results should be shown, saved, exported, or reviewed.
• Run python -X utf8 scripts/render_display_outputs.py --mode <excel|app|notes|codex> after a scan when the user chooses a display form.

Follow the operating rules

• Use exact dates. Say As of <date> instead of only today when reporting.
• Prefer official sources for rankings and journal scope. Do not invent rankings.
• Do not make ranking imports a hard prerequisite. The top-journal-family seed pool can run first and should remain trackable even when ranking data is missing.
• If a journal qualifies for full-track in any system, keep it in full-track.
• If a journal qualifies through the top-family seed pool, keep it in the watchlist and label the rank basis as top-family seed; ranking pending.
• If official ranking access is blocked, say so plainly and ask for an export or screenshot instead of guessing.
• If a paper is in the full-track pool but weakly related to the user's topic, keep it in the digest and mark it low-relevance rather than dropping it.
• Default to literature tracking outputs, not proposal-writing or review-writing outputs, unless the user explicitly asks.

Build or refresh the journal pool

• Confirm the scope with the user.
• Start with the top-journal-family seed pool from references/top-journal-families.md.
• Run scripts/build_top_family_watchlist.py when a concrete watchlist table, JSON, CSV, or source-id lookup is needed.
• Add field-specific or local journals when the user requests them.
• Save or restate the pool in a compact table when the user asks for the watchlist.

Establish the research profile

• Check for an existing profile in references/.
• Ask only for missing fields.
• Capture at least: core topics, secondary topics, methods, exclusion topics, and reading capacity.
• If the user has no fixed profile yet, build a draft from the last 5–10 papers they considered central.
• See references/example-profile.md for examples.

Track new papers

• Search the defined journal pool for new papers in the stated time window.
• Prefer scripts/scan_recent_papers.py for top-family seed scans.
• By default the scan script excludes news, briefings, corrections, and retractions; add --include-non-research only when explicitly asked.
• Record: title, year, authors, journal, publication date, DOI, pool source, topic tags, and relevance.
• Separate results into new, already seen, and needs manual verification.

Filter by topic

• Apply the research profile to title, abstract, keywords, methods, study system, and geography.
• Score relevance as high, medium, or low.
• In the relevant-only pool, keep only high and medium unless the user asks for exhaustive output.
• Distinguish core, proxy, and noise literature.

Generate the Monday report

• Group papers into must read, should read, watchlist, and notable methods or data.
• Highlight repeated trends across multiple papers.
• Include exact coverage dates for the report window.
• End with a short action list for the coming week.

Choose the display form

• The user decides the display form. If not specified, ask: How would you like the results displayed: Excel, interactive HTML table, Obsidian/Zotero notes, or inline preview?
• Supported forms: excel, app, notes, codex.
• Do not silently default to Markdown or HTML when the user has not chosen.

Maintain the reading list

• Track statuses: to-screen, queued, reading, noted, archived, skip.
• Store durable scan state in state/reading_state.json for recurring monitoring.
• For ad hoc previews, run scan scripts without --update-state.
• Promote high-relevance items first.
• Keep the queue small enough to fit the user's declared weekly capacity.

Keep the interaction practical

• Start by confirming the task mode: refresh pool, daily scan, topic filter, weekly report, reading backlog, or reminder list.
• For output tasks, confirm the display mode unless already specified.
• Ask one concise follow-up question only when missing information blocks a reliable answer.