keyword-redteam

Layer 4 of the keyword research pipeline. Spawns a "skeptical SEO" adversarial sub-agent to argue against every survivor of Layers 1-3. Catches mechanical-classifier blind spots — wrong SERP intent, h

SKILL.md

GitHub

Keyword Red-Team Skill

After Layers 1-3 mechanically vet the candidate pool, Layer 4 challenges the survivors with adversarial judgment. The mechanical layers see what's measurable; the redteam catches what's missable.

This is Ryan Law's "challenge / verify / push back" principle applied to keyword selection. Same shape as quality-check's adversarial sub-agent (a skeptical reader argues against the draft), just one stage earlier in the pipeline.

Input

/keyword-redteam

Reads:

  • content-pipeline/0-keywords/keyword-ideas.csv (after Layer 3 — must have bid_verdict, aio_verdict)
  • brand-config.md
  • .claude/skills/quality-check/SKILL.md lines 68-70 — the "agent says everything is fine" override pattern (we mirror it here)

Process

  1. Filter to survivors. Take rows where:

    • bid_verdict == PASS
    • aio_verdict ∈ {PASS, RISKY} (UNKNOWN is also kept — Layer 4 reasons about it)
    • Not yet in redteam_verdict column (idempotent: re-running doesn't re-redteam unchanged rows)
  2. Take the top 30 by current priority_score (after Layer 5 boosts will be applied — but Layer 4 runs before Layer 5 to keep the red-team pool small, since the redteam is the most token-expensive step).

    If fewer than 30 survivors exist, take all of them.

  3. Spawn the redteam sub-agent via Task with subagent_type=general-purpose and a self-contained brief. Per project memory ("Claude-in-Chrome work always runs on Sonnet 4.6") and the model-routing principle, this stays on the parent model (Opus by default) because the work is judgment-heavy:

    You are a senior SEO consultant who has reviewed 1,000+ content briefs. You are reviewing 30 candidate keywords for {brand} — see {brand-config-path}. The brand has already passed each candidate through:

    • Layer 1: keyword discovery (competitor gap + seed/modifier expansion + AI-search citation gap)
    • Layer 2: BID method — Business potential (brand_fit, product_fit), Intent (SERP-classified), Difficulty (KD% + AS)
    • Layer 3: AI Overview cannibalization check (0-10 completeness scoring on AIOs that exist)

    Each candidate's row also exposes Semrush-native columns worth reasoning about: an intents array (Semrush's per-keyword intent classifier — multi-label across informational / navigational / commercial / transactional) and a cluster_id from Keyword Strategy Builder (the topical cluster the keyword belongs to). Use both as you work through (a)-(d): intents is a higher-quality intent signal than the URL-pattern heuristic, and cluster_id tells you whether neighbouring keywords in the same cluster already rank for the brand or for a single dominant competitor.

    Your job: argue against each candidate. For each keyword, address all four:

    (a) SERP-intent classification challenge. Layer 2 classified the SERP intent from URL patterns and titles. Why might that classification be wrong for this specific keyword? Is the SERP a hybrid where the dominant intent depends on user context? Would a hands-on reader of the SERP disagree with the classifier?

    (b) AIO trajectory. Even if there's no AIO today, will Google add one in 12 months? What's the shape of the query — definitional, comparative, walkthrough, opinion-driven, news-driven? Definitional and walkthrough queries are getting AIOs fast; opinion / comparison / fresh-news queries are AIO-resistant. Is this candidate AIO-bait?

    (c) Hidden difficulty. Layer 2 used AS + KD% as proxies. What's hiding in the link graph that those metrics miss? E.g.: top result is a publisher with editorial link velocity competitors can't match; SERP is dominated by a single industry institution; YouTube videos rank in the SERP and crowd out blog posts; SERP has high domain diversity (suggesting it's hard to displace any one player). Cross-reference cluster_id: are neighbouring keywords in the same cluster also dominated by the same authority sites? If so, the difficulty is cluster-deep, not query-specific.

    (d) Business potential overstatement. Layer 2 used brand_fit + product_fit. Is this keyword a "vanity rank" that produces traffic but no users / revenue / brand mentions? Would shipping content for this keyword actually move the brand's KPIs, or is it a number-gone-up trap?

    For EACH of the 30 candidates, end with one of three verdicts:

    • KEEP — the candidate is genuinely good; proceed to writing pipeline.
    • DROP — fatal flaw in (a)-(d) that the mechanical layers missed; remove from queue.
    • REVISE_PRIORITY — keep but adjust priority_score by +X.X or -X.X (capped at ±2.0). Explain why the score is wrong.

    Be specific. Don't say "this might be hard." Say "the top three results are all from Healthline, Mayo Clinic, and Cleveland Clinic — link velocity from medical institutions is editorial-graded; brand DR + 30 wouldn't beat them."

    Don't be diplomatic. If 25 of the 30 candidates are good, KEEP 25 and DROP 5. If only 5 are good, that's the verdict — KEEP 5 and DROP 25. Don't pad KEEP rates to look productive.

    Output format (one block per candidate, in the order I give you):

    ## {n}. {keyword}
    Verdict: {KEEP | DROP | REVISE_PRIORITY +X.X | REVISE_PRIORITY -X.X}
    (a) SERP-intent challenge: <one sentence>
    (b) AIO trajectory: <one sentence>
    (c) Hidden difficulty: <one sentence>
    (d) Business potential check: <one sentence>
    One-line summary: <why your verdict>
    

    Provide the agent with:

    • The 30 candidates with their full vetting metadata as a markdown table
    • The brand-config path
    • The output format template
  4. Validate the response.

    • Every candidate has a verdict line
    • Verdict is one of: KEEP, DROP, REVISE_PRIORITY +X.X, REVISE_PRIORITY -X.X (X.X within ±2.0)
    • Each (a)-(d) is filled and specific (not generic placeholder)
    • KEEP rate sanity check: if > 80% of verdicts are KEEP AND total response < 200 words per candidate average, treat as a low-quality "rubber-stamp" pass and re-run the agent with this addendum:

      Your previous response had a KEEP rate of {X}% across {N} candidates with weak per-candidate critique. Re-do the redteam with these constraints: (1) maximum KEEP rate of 70% — at least 30% of candidates must DROP or REVISE_PRIORITY-down; (2) per-candidate critique must be at least 60 words and name something specific (a competitor domain, an SERP feature, a brand-fit gap, a difficulty metric, etc.); (3) if you genuinely believe most candidates are excellent, prove it by producing the strongest available counter-argument anyway and rebutting it. This mirrors quality-check/SKILL.md lines 68-70 — same override on adversarial-agent rubber-stamping.

  5. Apply verdicts to the CSV:

    • DROP → redteam_verdict=DROP, the row is excluded from keyword-queue.csv
    • KEEP → redteam_verdict=KEEP, no priority adjustment
    • REVISE_PRIORITY → redteam_verdict=REVISE_PRIORITY, redteam_priority_delta set to the proposed delta (capped at ±2.0)
    • Each row also gets redteam_critique_summary (the one-line summary from the agent)
  6. Write the full critique to content-pipeline/0-keywords/redteam-notes.md (preserves the (a)-(d) detail per candidate for editor / future reference).

  7. Print summary:

    Redteam over N candidates:
      KEEP:              X
      DROP:              X
      REVISE_PRIORITY+:  X (sum delta: +X.X)
      REVISE_PRIORITY-:  X (sum delta: -X.X)
    Top 3 DROPs:
      - <keyword> — <one-line summary>
    Top 3 REVISE-:
      - <keyword> — <one-line summary> (delta: -X.X)
    

Output

Updated content-pipeline/0-keywords/keyword-ideas.csv with redteam_verdict, redteam_priority_delta, redteam_critique_summary.

Full critique at content-pipeline/0-keywords/redteam-notes.md.

Quality checklist

  • Every candidate in the top-30 input has a verdict (no skipped rows)
  • At least 1 DROP verdict (gate is doing real work — if all KEEP, run the override)
  • Per-candidate (a)-(d) critique is specific, not generic
  • redteam_priority_delta values are within ±2.0
  • redteam-notes.md preserves the full agent output for future review

Why this layer matters

The mechanical layers (BID, AIO) are necessary but not sufficient. Real keyword selection is judgment-heavy: a high-DR competitor isn't always unbeatable (they may have stale content); a low-AIO query may sprout one tomorrow; a "great brand fit" keyword may be a vanity rank in disguise. Layer 4 is the only layer that reasons about the future and the implicit. Without it, the autonomous pipeline ships content for keywords that look right on paper but fail in the SERP.

When the redteam wants to KEEP everything

Per the override protocol above. The brief is structured to force a contrarian read; if the agent comes back with all-praise, the brief failed to bite. Re-run with the explicit "max 70% KEEP rate, 60-word minimum critique" addendum.

When the redteam wants to DROP everything

Rare. Usually means Layers 1-3 produced a genuinely poor pool (e.g., gap analysis returned only branded competitor terms; seeds were too generic). Don't disagree with the agent — if it has specific (a)-(d) reasoning per candidate, the pool IS bad. Halt the pipeline, log to content-pipeline/0-keywords/cache/redteam-poor-pool.log, and signal to the keyword-research-pipeline orchestrator that Layer 1 needs investigation. Don't ship a queue from a poor pool.

Idempotency

Re-running this skill on a CSV where some rows already have redteam_verdict skips those rows. The agent only sees rows with a missing verdict. This means:

  • A failed run can be resumed by re-invoking the skill
  • Adding new candidates from a fresh /content-gap-analysis doesn't re-redteam old candidates

To force a re-redteam (e.g., after threshold tuning or a major brand-config change), pass --regen to clear the redteam columns first.

Cost notes

Layer 4 is the most expensive layer per keyword (one Opus/Sonnet call per 30-candidate batch). Per-run budget: ~30K tokens output + 10K tokens input ≈ $0.40-0.60 per keyword-research-pipeline invocation. The orchestrator's weekly cadence keeps this annual cost under $30 — cheap insurance against shipping a single dead-traffic article.