Research Adversarial

Research Adversarial Skill

Phase 3 of PLEAA-392 (PLEAA-418, 2026-05-06). The research dossier is the single source of truth that flows into outline → draft → cited draft. If the dossier missed the strongest competitor angle or didn't cite a load-bearing claim, every downstream stage inherits that flaw.

The orchestrator dispatches this skill after stage 1 (/research) writes content-pipeline/1-research/{slug}.md and before stage 3 (/outline) reads it. Stage 2 (/brand-reference) is independent and can run in parallel either way. If the adversarial verdict is FAIL and budget remains, the orchestrator re-dispatches /research with the critique, then re-runs this skill.

Input

For slug {slug}:

• content-pipeline/1-research/{slug}.md (the dossier to push back on)
• content-pipeline/1-research/{slug}-deep.md (if present — the deep-research appendix from OpenRouter)
• content-pipeline/1-research/{slug}-data.json (chartable numbers — do they match what the prose dossier claims?)
• content-pipeline/0-context/{slug}.md (if present — Neo's context string)
• brand-config.md (audience + voice)

Process

1. Spawn a Task sub-agent with the brief below.
2. Combine into content-pipeline/quality-checks/{slug}-research-adversarial.md in the format under Output.
3. Print the verdict + CRITICAL count.

Adversarial sub-agent brief

You are a skeptical research lead reviewing the dossier at content-pipeline/1-research/{slug}.md before it feeds an outline. The writer hasn't seen this dossier yet — your job is to make sure they don't build on a weak foundation.

Read in order:

1. content-pipeline/1-research/{slug}.md
2. content-pipeline/1-research/{slug}-deep.md (if it exists)
3. content-pipeline/1-research/{slug}-data.json (the chartable numbers)
4. content-pipeline/0-context/{slug}.md (if it exists — Neo's context)
5. brand-config.md

Push back on five questions. Be specific — quote the dossier's own sentences, point at section headers, name URLs:

1. Citations. Does every numerical claim and every surprising statement cite a source URL or an internal Semrush/Strapi reference? List up to 3 uncited load-bearing claims by quoting them.
2. Surprising findings. The dossier should surface 3 surprising findings. Are they actually surprising to a reader who already knows this niche, or are they table-stakes facts dressed up as insights? Name each "surprise" and rate it actually-surprising or dressed-up-table-stakes.
3. Strongest competitor angle missed. Look at the SERP top-5 the dossier summarized. Which competitor angle is the strongest, and does the dossier capture how to beat it — or just acknowledge it exists? Name the angle and the URL.
4. Data consistency. Open {slug}-data.json and spot-check 3 keys against the prose dossier. Do the numbers match? Any key in the JSON the prose never mentions, or any prose number not in the JSON?
5. Brand fit. Does the dossier surface enough material the brand can own (Pleasur.AI-specific data, internal modules from 2-reference/{slug}.md, the 4 product pillars from brand-config) — or is it a generic survey?

Output:

• 5–8 specific findings tagged CRITICAL, HIGH, MEDIUM, LOW.
  • CRITICAL = drafting on this will produce a citation-thin or wrong-angle article. Examples: load-bearing stat with no source, a dressed-up "surprise" that's actually a 5-year-old fact, missing the strongest competitor angle entirely, JSON keys that contradict the prose.
  • HIGH = the dossier is correct but thin in a way that will hurt the draft.
• 1 thing that works.
• A one-line Verdict: PASS or Verdict: FAIL.

Do NOT rewrite the dossier or suggest new sources. Under 600 words.

Output

Write content-pipeline/quality-checks/{slug}-research-adversarial.md with the same skeleton as the outline-adversarial skill. The ## Verdict: line MUST be exactly ## Verdict: **PASS** or ## Verdict: **FAIL**.

How the orchestrator handles FAIL

Same loop as outline-adversarial, but using stage key research:

1. python scripts/adversarial_runlog.py can-revise {slug} research
2. If exit 0 (budget remains): increment, re-dispatch /research with the adversarial CRITICAL list as a revision brief, then re-run this skill.
3. If exit 1 (exhausted): write content-pipeline/9-needs-review/{slug}.md and HALT.

Why this skill exists

The dossier is upstream of every other stage. A bad dossier produces a bad outline produces a bad draft. Editing prose downstream can't recover from "this article was built on the wrong angle." Catch the angle problem here.

When the adversarial sub-agent says everything is fine

Same calibration rule as outline-adversarial: re-run with a stronger brief if the critique has fewer than 3 specific findings.