Keyword Discovery Skill

Walk the user through a complete keyword research pipeline, producing a prioritized keyword plan they can act on immediately. Assume the user has never done keyword research before — explain every concept in plain language, but keep the process moving.

Cross-model note

Works identically on current Opus / Sonnet / Haiku models. Expansion + intent classification + clustering all benefit from batch parallelism where per-keyword work is independent.

Execution: parallel-first

• Phase 1 seed gathering: user interview serializes naturally
• Phase 2 expansion: long-tail / question / modifier / LSI / medium-tail expansion runs per-seed; per-seed WebSearch-autocomplete checks parallelize
• Phase 3 intent classification: per-keyword SERP inspection independent — batch WebSearch calls
• Phase 4 difficulty assessment: per-keyword SERP + competitor inspection independent — batch
• Phase 5 clustering: synthesis step, not parallelizable
• Phase 6 prioritization: synthesis step, not parallelizable

Capability Honesty

Read references/CAPABILITIES.md. Key constraints:

• Real search volume numbers (e.g., "1,200 searches/month") require a keyword-data API — Path B
• Real keyword difficulty scores (e.g., "KD 42") require link-graph + SERP competitor strength — Path B
• What this skill DOES provide: observable SERP signals (who ranks, SERP feature presence, autocomplete prominence, competition density) interpreted qualitatively (High/Medium/Low, Easy/Moderate/Hard)

When the user wants real numbers, recommend a DataForSEO / Ahrefs / Semrush MCP integration.

Mode Flags

• --non-interactive — Use reasonable defaults for Phase 1 questions, record unanswered items in awaiting_user_review

Overview

Six phases in order:

1. Seed Keyword Discovery — gather raw inputs
2. Keyword Expansion — broaden the list
3. Search Intent Classification — tag every keyword
4. Qualitative Difficulty Assessment — observable competition signals
5. Topic Cluster Mapping — organize into hierarchy
6. Prioritization — produce ranked action plan

Work through each phase sequentially. Present findings after each phase before moving on.

Phase 1: Seed Keyword Discovery

Gather Context

Before generating a single keyword, ask:

1. Business basics — What does the business do? Products / services?
2. Target audience — Ideal customers? Language / jargon they use?
3. Top 3 competitors — Up to three main rivals (URLs or brands)
4. Existing keywords — Already targeting any? List them, or "none"

Seed Sources

Mine seeds from every source below:

Deliver as a numbered table. Aim for 30-50 seed keywords before moving on.

Phase 2: Keyword Expansion

Expand each seed using four strategies. Go from ~40 seeds to 150-300 candidate keywords.

2a. Long-Tail Variations (3-5 word phrases)

Add specificity. Long-tail keywords convert better because they match precise needs.

2b. Question-Based Keywords

Prepend: who, what, when, where, why, how.

2c. Modifier Keywords

• Best / Top / Review / Comparison / vs — commercial intent
• Alternative / Cheap / Free — budget-conscious buyers
• Near me / [City] — local intent
• Template / Example / Checklist — resource seekers
• For beginners / For enterprises / For [audience] — audience fit

2d. LSI / Semantic Keywords

Related terms that appear alongside the seed in top-ranking content.

2e. Medium-Tail Keywords

"Sweet spot" — 2-3 words, moderate competition. Often fastest wins.

Group the expanded list by strategy so the user can see where each keyword came from.

Phase 3: Search Intent Classification

Every keyword gets exactly one intent tag.

Intent Types

How to Classify

Look at the actual SERP (WebSearch):

• Shopping carousel / product ads → Transactional
• Video pack or PAA → likely Informational
• Comparison articles dominate → Commercial
• Single brand dominates positions 1-3 → Navigational

Mixed intent → label dominant, note secondary.

Add an "Intent" column to the keyword table.

Phase 4: Qualitative Difficulty Assessment

Honest framing: This phase uses observable SERP signals to estimate relative difficulty (Easy / Moderate / Hard). It does NOT produce numerical difficulty scores. For real Keyword Difficulty metrics, integrate a DataForSEO / Ahrefs / Semrush MCP (Path B).

Factors to Evaluate (per target keyword, batch WebSearch)

Difficulty Scale (qualitative)

• Easy: Thin/outdated content ranks; few strong domains; featured snippet available. Well-written article can rank in weeks.
• Moderate: Mix of strong and weak pages; requires comprehensive content + some backlinks. Expect 2-6 months.
• Hard: P1 is all high-authority with deep content. Requires significant authority, backlinks, time (6-12+ months).

Add "Difficulty" and "Observed Opportunity" columns:

#  | Keyword | Intent | Difficulty | Observed Opportunity
1  | project management for nonprofits | Commercial | Easy | Featured snippet slot available, weak top-3
2  | project management software | Commercial | Hard | Heavy ad spend, strong DR domains

Volume Signal (qualitative, not numerical)

Use observable signals to estimate relative volume:

• High signal: prominent autocomplete, multiple SERP features, heavy ad coverage
• Medium signal: autocomplete present, some SERP features
• Low signal: no autocomplete, thin SERP, no ads

Do NOT invent "monthly search volume" numbers. Mark relative volume as High/Medium/Low.

Phase 5: Topic Cluster Mapping

Organize keywords into a content hierarchy.

Hierarchy

Pillar Topic (broad, 1-2 words, high relative volume)
  +-- Subtopic A (medium-tail, 2-3 words)
  |     +-- Long-tail query 1
  |     +-- Long-tail query 2
  +-- Subtopic B
        +-- Long-tail query 3
        +-- Long-tail query 4

How to Build Clusters

1. Identify pillars — broadest, highest-relative-volume terms → comprehensive pillar pages (2,000-4,000 words)
2. Group subtopics — medium-tail as subtopics → supporting articles (1,000-2,000 words)
3. Attach long-tail — become sections within subtopic articles, standalone FAQ entries, or short posts
4. Map internal links — every subtopic links up to pillar; pillar links down to every subtopic; subtopics in same cluster link to each other where relevant

Present the full cluster map and confirm with the user before prioritization.

Phase 6: Prioritization

Scoring Criteria (qualitative)

The Three Lists

#### 1. Top 10 Quick Wins Keywords where content already exists (or is easy to create), competition signals easy, rank in weeks.

#### 2. Top 10 Strategic Targets High-relevance with moderate competition. Dedicated content + some link building. Expect results in 2-6 months.

#### 3. Top 10 Long-Term Plays High-competition, high-reward. Build toward these by first ranking for easier related keywords.

Content Type Recommendations

The 70/20/10 Content Mix

• 70% Foundation — how-to guides, tutorials, listicles, FAQ. Safe, proven formats. Start here.
• 20% Growth — original research, data-driven posts, expert roundups, detailed comparisons. Higher effort, differentiates.
• 10% Moonshots — interactive tools, contrarian pieces, original frameworks, viral-potential content. Most underperform but wins are outsized.

Tag each content calendar entry with its mix category.

Delivery Checklist

• [ ] 30-50 seed keywords with sources
• [ ] 150-300 expanded keywords grouped by strategy
• [ ] Every keyword tagged with search intent
• [ ] Top 50+ keywords assessed for qualitative difficulty
• [ ] Complete topic cluster map with pillar-subtopic hierarchy
• [ ] Internal linking plan between cluster pages
• [ ] Top 10 Quick Wins with specific actions
• [ ] Top 10 Strategic Targets with content recommendations
• [ ] Top 10 Long-Term Plays with cluster-building roadmap
• [ ] Content type recommendation for each priority keyword
• [ ] 70/20/10 mix labels on all recommended content
• [ ] Measurement-gaps note: "For numerical search volumes and KD scores, integrate a DataForSEO / Ahrefs / Semrush MCP"

Tone and Style

• Plain language. Define jargon first time used.
• Specific actions. "Write a blog post" is not actionable. "Write a 2,000-word comparison of X vs Y targeting commercial intent" is.
• Tables for any list > 5 items.
• Summary after each phase.
• When lacking real search volume data, say so honestly and use relative estimates based on observable signals.