Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when the user asks to "find keywords", "挖词", or "搜什么词"; prioritizes search volume, keyword difficulty, intent, and topic clusters from provided or connected data. Not for competitor-relative coverage gaps — use content-gap-analysis. 关键词研究/内容选题
.claude/skills/aaron-he-zhu-keyword-research/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | 52% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 90% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 56% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 68% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 77% | 0% |
Discovers, scores, and clusters keywords for SEO and GEO planning.
Research keywords for [topic/product/service]What keywords is [competitor URL] ranking for that I should target?Expected output: a prioritized keyword brief plus the standard handoff summary for memory/research/.
memory/hot-cache.md, memory/open-loops.md, and memory/research/.Unknown with its gap reason, never N/A; keywords are grouped into pillar + cluster hubs; and the deliverable names at least 3 prioritized Quick Win / Growth / GEO opportunities.> Emit the standard shape from skill-contract.md §Handoff Summary Format.
Optional integrations: ~~SEO tool, ~~search console. Without tools, ask for seed keywords, audience, goals, and any known metrics. See CONNECTORS.md.
Zero-dependency local helper (no tool needed): python3 "${CLAUDE_PLUGIN_ROOT}/scripts/connectors/suggest.py" "<seed>" --expand harvests free keyword ideas from Google Autocomplete (⚠️ unofficial endpoint). Search volume / difficulty still needs ~~SEO tool or own Search Console data. See scripts/connectors/README.md.
Keyless live-SERP sampling: python3 "${CLAUDE_PLUGIN_ROOT}/scripts/connectors/firecrawl.py" search "<candidate keyword>" --limit 10 (Firecrawl keyless free tier, ~1,000 credits/mo, no key needed) shows who actually ranks for a candidate — feed the top-10 domains and formats into the intent check and the difficulty read as Measured evidence instead of guessing. Volume still needs ~~SEO tool or GSC.
Keyless topic-demand proxy: python3 "${CLAUDE_PLUGIN_ROOT}/scripts/connectors/pageviews.py" "<Topic_Article>" --months 12 returns a topic's real Wikipedia-attention series — Measured direction and seasonality evidence when no volume tool is connected. It is attention, not search volume: use it to rank topics against each other and time them, never to quote a volume number.
Striking-distance shortcut (when ~~search console is connected): before broad discovery, mine your own GSC query data for terms already ranking in positions ~5–20 — page-one tail and page two. These are proven demand a small push can convert, so they are the fastest opportunity set. The Search Analytics API sorts by clicks and has no position filter, so request a high rowLimit and filter the 5–20 window client-side, then attach volume / difficulty / intent to that shortlist. Work this set first; treat its metrics as Measured.
When a user requests keyword research, run eight phases and announce each as [Phase X/8: Name]:
Opportunity = (Volume × Intent Value) / Difficulty, with Intent Value 1 / 1 / 2 / 3.Label every metric Measured (tool/export), User-provided, Calculated, Estimated, Proxy, or Unknown; retain query, locale, language, source ref, observation time, and window per field. Preserve conflicting sources. If an applicable metric is unavailable, mark it Unknown with a missing reason — N/A is only for a genuinely non-applicable field. An attention proxy never becomes search volume, and an Unknown decision-critical input makes the opportunity score NOT_SCORED.
When you have richer signals than volume/difficulty alone, add a second pass on top of the Opportunity score:
Tag each keyword by funnel stage from its pattern:
Work BOFU first when revenue is the goal; use TOFU/MOFU for reach and GEO answer coverage. (Impact×Confidence + funnel-stage scoring adapted from an external SEO-ops competitive analysis.)
Quality bar: every recommendation includes at least one specific number. Rewrite generic advice into a concrete keyword + volume + difficulty + reason.
> Reference: See references/instructions-detail.md for the full 8-phase templates, expansion patterns, intent table, difficulty tiers, opportunity matrix, GEO indicators, cluster template, actionable-vs-generic examples, and advanced usage.
See references/example-report.md for a full worked sample.
Write path: memory/research/keyword-research/YYYY-MM-DD-<topic>.md; promote durable keyword priorities to memory/hot-cache.md. See Skill Contract §Save Results Template.
Primary: competitor-analysis. Also: content-gap-analysis and serp-analysis.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→fail | 40,801 | 18,768 | -54% | 1 | 1 | 0% | 6,412 | 2,538 | -60% | 0 | 0 | — |
case-19 | pass→pass | 26,149 | 20,161 | -23% | 1 | 1 | 0% | 3,195 | 4,123 | +29% | 0 | 0 | — |
case-12 | fail→pass | 15,330 | 10,803 | -30% | 1 | 1 | 0% | 1,725 | 2,614 | +52% | 0 | 0 | — |
case-01 | fail→fail | 37,509 | 30,720 | -18% | 1 | 1 | 0% | 5,721 | 6,955 | +22% | 0 | 0 | — |
case-03 | fail→fail | 33,710 | 40,975 | +22% | 1 | 1 | 0% | 6,299 | 9,231 | +47% | 0 | 0 | — |
case-04 | pass→pass | 14,116 | 14,018 | -1% | 1 | 1 | 0% | 1,674 | 3,539 | +111% | 0 | 0 | — |
case-05 | pass→pass | 20,698 | 25,886 | +25% | 1 | 1 | 0% | 2,686 | 5,360 | +100% | 0 | 0 | — |
case-06 | fail→pass | 19,517 | 16,998 | -13% | 1 | 1 | 0% | 2,069 | 3,931 | +90% | 0 | 0 | — |
case-13 | fail→pass | 15,406 | 9,607 | -38% | 1 | 1 | 0% | 1,595 | 2,495 | +56% | 0 | 0 | — |
case-07 | fail→pass | 18,236 | 19,416 | +6% | 1 | 1 | 0% | 2,502 | 4,207 | +68% | 0 | 0 | — |
case-08 | pass→pass | 16,097 | 11,466 | -29% | 1 | 1 | 0% | 2,070 | 2,910 | +41% | 0 | 0 | — |
case-09 | fail→pass | 15,944 | 13,876 | -13% | 1 | 1 | 0% | 1,783 | 3,156 | +77% | 0 | 0 | — |
case-10 | fail→pass | 20,480 | 12,973 | -37% | 1 | 1 | 0% | 2,441 | 3,119 | +28% | 0 | 0 | — |
case-11 | pass→pass | 15,713 | 13,229 | -16% | 1 | 1 | 0% | 1,764 | 3,146 | +78% | 0 | 0 | — |
case-14 | pass→pass | 15,664 | 10,912 | -30% | 1 | 1 | 0% | 1,781 | 2,759 | +55% | 0 | 0 | — |
case-15 | pass→pass | 10,019 | 9,038 | -10% | 1 | 1 | 0% | 801 | 2,519 | +214% | 0 | 0 | — |
case-16 | fail→pass | 23,380 | 35,352 | +51% | 1 | 1 | 0% | 3,032 | 7,656 | +153% | 0 | 0 | — |
case-17 | fail→pass | 13,276 | 7,170 | -46% | 1 | 1 | 0% | 1,290 | 2,125 | +65% | 0 | 0 | — |
case-18 | fail→pass | 17,653 | 8,525 | -52% | 1 | 1 | 0% | 2,076 | 2,340 | +13% | 0 | 0 | — |
case-20 | fail→pass | 14,802 | 13,235 | -11% | 1 | 1 | 0% | 1,521 | 3,062 | +101% | 0 | 0 | — |
case-21 | fail→pass | 20,676 | 17,939 | -13% | 1 | 1 | 0% | 2,693 | 4,083 | +52% | 0 | 0 | — |
case-22 | pass→pass | 14,244 | 7,675 | -46% | 1 | 1 | 0% | 1,429 | 2,187 | +53% | 0 | 0 | — |
case-23 | fail→pass | 14,684 | 9,349 | -36% | 1 | 1 | 0% | 1,758 | 2,598 | +48% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 22 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +52 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/13/2026 | +55% |
Other measured skills in the registry, with their headline benchmark lift.