Internet Skill Radar

Find decision-worthy Skills, not merely popular ones. Treat every marketplace metric as an unverified discovery signal until the underlying repository and SKILL.md are inspected.

Set the scan

Infer missing settings unless they materially change the result:

• Goal: adopt, learn, build, invest, benchmark, or monitor.
• Audience: the user's role, recurring tasks, stack, and risk tolerance.
• Window: default to 30 days for discovery and compare against the previous snapshot when available.
• Breadth: default to 20 verified candidates and a final shortlist of 5.
• Compatibility: accept portable SKILL.md Skills; flag product-specific dependencies.

For current scans, browse the internet. Record the scan time, query, source URL, source tier, and observed metric for every candidate.

Discover broadly

Read references/sources.md before selecting sources.

1. Search Tier 1 official catalogs first.
2. Search at least two independent directories for breadth.
3. Search GitHub directly for recent or fast-changing SKILL.md files.
4. Use community posts, papers, and trend pages only as leads.
5. Add candidates from the user's URLs or watchlist.

Do not let one catalog dominate the shortlist. Canonicalize each candidate as owner/repo/path, collapse mirrors and forks, and retain the original author when discoverable.

Verify before scoring

Open the canonical repository and inspect:

• The actual SKILL.md, not only its marketplace description.
• Trigger specificity, workflow completeness, reusable resources, tests or examples.
• Commit recency, meaningful contributors, issue health, license, release or tag history.
• Scripts, dependencies, install steps, requested permissions, network calls, secret access, and downloaded executables.
• Consistency between the listing, SKILL.md, scripts, and repository purpose.

Never install or execute a candidate during a scan unless the user explicitly asks. Quarantine candidates with hidden instructions, obfuscation, credential harvesting, destructive commands, unexplained remote execution, privilege escalation, or repo/Skill mismatch.

Score consistently

Read references/methodology.md. Score each 0-100 dimension from cited evidence. Use scripts/score_skills.py when ranking more than three candidates.

Keep three values distinct:

• opportunity: usefulness before uncertainty and risk.
• confidence: evidence completeness and independence.
• risk: potential downside; this is not canceled by popularity.

Keep GitHub and X signals visible and separate:

• GitHub: lifetime Stars plus Star growth between snapshots.
• X: recent seven-day mentions, public engagement, and independent authors.
• Use null, not zero, when a social signal is unavailable.

Use the calculated value_score for ordering, then apply hard safety gates. Never invent missing metrics; mark them unknown and lower confidence.

Compare with history

When a previous snapshot exists, match canonical IDs and calculate:

• New entrants and removals.
• Rank, adoption, activity, and value-score changes.
• SKILL.md content-fingerprint changes, plus material permission, ownership, or license changes.
• Staleness: no meaningful update despite unresolved compatibility or security issues.

If no snapshot exists, label the result baseline; do not call a Skill “rising” from a single observation.
Keep at most 30 observation points per Skill. Treat a missing prior content fingerprint as version-baseline migration, not as an update.

Report the radar

Lead with the top decisions, then provide:

Rank	Skill	What it unlocks	Value	Confidence	Risk	Momentum	Action

For each shortlisted Skill, include:

• Canonical repository and direct SKILL.md link.
• One-sentence use case tied to the user's work.
• Evidence for value and momentum.
• Important permissions, dependencies, and caveats.
• Action: adopt, sandbox-test, watch, build-alternative, ignore, or quarantine.

End with:

• Build gaps: high-demand needs with weak or duplicated current offerings.
• Watch changes: exact signals to check next time.
• Method note: scan date, window, sources, missing data, and whether this is a baseline.

Use calibrated language. Separate facts, inferences, and recommendations. Cite live sources near the claims they support.

Use the web dashboard

Use assets/dashboard/index.html when the user asks for a visual, web, dashboard, or radar view.

For the connected mode on Windows:

1. Tell the user to double-click assets/dashboard/启动联网版.cmd.
2. Open http://127.0.0.1:8765/; do not use file:// for connected features.
3. Use 立即联网更新 for an on-demand scan.
4. Use 搜索 GitHub for an explicit task keyword; inspect the returned SKILL.md candidates before adding them to the radar.
5. Use 启用每日任务 to create an explicit Windows daily task at 08:00.
6. Open a candidate to read the Chinese explanation, evidence, risk, and permissions.
7. Use 安装到 Codex or 安装到 Hermes only after reviewing and confirming the install dialog.
8. Use 导出通用 Skill 包 when the target Agent is not Codex or Hermes; keep the complete folder because bundled resources may be required.

The local service binds only to 127.0.0.1. It scans official catalogs and current GitHub repositories, keeps rolling history, and rejects overwrite, path traversal, oversized packages, symbolic links, and quarantined candidates. Codex installs go to $CODEX_HOME/skills or ~/.codex/skills. Hermes installs prefer the official hermes skills install command and otherwise use $HERMES_HOME/skills/radar or ~/.hermes/skills/radar. Portable exports retain the original files and add a manifest plus Chinese installation notes; compatibility labels are signals, not proof of full runtime support.

For static mode, open assets/dashboard/index.html directly and import a snapshot through 导入 JSON. Static mode cannot scan, schedule, or install.

Save monitoring state

Only save files when requested or when a recurring scan needs continuity. Use:

• skill-radar-snapshot.json for machine-readable candidates, evidence, scores, and scan metadata.
• skill-radar-report.md for the human report.

Preserve previous snapshots or add a timestamp instead of overwriting the only baseline.