Outline Adversarial

Outline Adversarial Skill

Phase 3 of PLEAA-392 (PLEAA-418, 2026-05-06). Neo's brief:

There's something I really like — pushback where the blog says something and a skill basically pushed back, and then he can make multiple drafts. It's like a thing kind of pushed back, and maybe if something is wrong, he can rethink it. I like that.

The orchestrator dispatches this skill after stage 3 (/outline) writes content-pipeline/3-outlines/{slug}.md and before stage 4 (/product-mentions) annotates it. If the adversarial verdict is FAIL and a revision budget remains, the orchestrator re-dispatches /outline with the critique as a brief, then re-runs this skill. The runlog at .runs/{slug}-budgets.json keeps the count.

Input

For slug {slug}:

• content-pipeline/3-outlines/{slug}.md (the outline to push back on)
• content-pipeline/1-research/{slug}.md (so the adversarial knows what the SERP top-5 covered and what the angle is supposed to be)
• content-pipeline/2-reference/{slug}.md (so the adversarial knows what internal modules already exist; outlines that re-explain owned material lose)
• templates/editorial-principles-visuals.md (the 9-step "does this section earn a visual?" rule)
• brand-config.md (audience, voice, forbidden patterns)

Process

1. Spawn a Task sub-agent with the brief below. The agent reads the inputs and returns a critique. Do NOT let the agent rewrite the outline — its only job is to push back.

2. Combine the agent's output into content-pipeline/quality-checks/{slug}-outline-adversarial.md with the exact format described under Output.

3. Print the verdict + CRITICAL count to stdout so the orchestrator can read it in the agent return.

Adversarial sub-agent brief

You are a skeptical content editor pushing back on the outline at content-pipeline/3-outlines/{slug}.md. The article hasn't been drafted yet — your only job is to make sure the structure is sound BEFORE someone burns 2,000 words on a bad spine.

Read these in order:

1. content-pipeline/3-outlines/{slug}.md
2. content-pipeline/1-research/{slug}.md (note the SERP top-5 + the recommended angle)
3. content-pipeline/2-reference/{slug}.md (existing brand articles)
4. templates/editorial-principles-visuals.md
5. brand-config.md

Push back on five questions. Be specific — point at H2s by name and line ranges, not vague impressions:

1. MECE. Are the H2s mutually exclusive and collectively exhaustive? Any overlap between sections? Any obvious gap a competing SERP article covers that this outline doesn't?
2. BLUF on every opener. Does each H2 stub start with a clear "answer first" sentence, or do any of them throat-clear ("In today's world…")?
3. Problem-agitate-solution arc. Does the outline take the reader from pain to resolution, or is it a flat list of facts?
4. Differentiation vs SERP top-5. Does this outline cover the topic better than the SERP top-5 the research dossier surfaced, or just match them? Name at least one angle a competitor takes that this outline duplicates and one angle this outline takes that competitors miss.
5. Visuals earn their place. For each section with a non-none Visual, does the section actually need one per the 9-step rule, or is it a decorative placeholder? Which sections without a visual should have one?

Output:

• 5–8 specific findings, each tagged CRITICAL, HIGH, MEDIUM, or LOW.
  • CRITICAL = drafting on top of this will produce a worse article than the SERP top-5 or violate brand-config rules. Examples: missing H2 the SERP all cover, two H2s that say the same thing, BLUF that contradicts the visual, comparison section without honest verdicts on Pleasur.AI.
  • HIGH = the article will pass quality gates but feel weak.
  • MEDIUM/LOW = polish.
• 1 thing that genuinely works (so you stay calibrated, not contrarian for its own sake).
• A one-line Verdict: PASS or Verdict: FAIL. PASS = no CRITICAL findings. FAIL = at least one CRITICAL.

Do NOT rewrite the outline, suggest replacement H2s, or be polite. Under 600 words.

Output

Write content-pipeline/quality-checks/{slug}-outline-adversarial.md:

# Outline Adversarial — {slug} (Pass {N})

## Verdict: **PASS** or **FAIL**

Pass {N} of {budget+1} (revision budget BLOG_AGENT_OUTLINE_REVISION_BUDGET={budget}).

## Findings

### CRITICAL

- [section/line] one-line finding. Why it kills the draft.

### HIGH

- ...

### MEDIUM

- ...

### LOW

- ...

## What Works

- [section/line] one specific thing that's good.

## Recommendation

- If Verdict is PASS: advance to stage 4 (/product-mentions).
- If Verdict is FAIL: re-dispatch /outline with the CRITICAL items as the
  revision brief (orchestrator handles this), then re-run this skill.

The ## Verdict: line MUST be exactly that format (markdown bold) so pipeline_gate.py can parse it.

How the orchestrator handles FAIL

The orchestrator (/blog-pipeline) calls this skill after stage 3 and reads the verdict. On FAIL:

1. Read .runs/{slug}-budgets.json (or call python scripts/adversarial_runlog.py used {slug} outline).
2. If used < budget:
   • Increment via python scripts/adversarial_runlog.py increment {slug} outline.
   • Re-dispatch stage 3 (/outline) with the adversarial critique as the revision brief — specifically the CRITICAL findings.
   • Re-run this skill.
3. If used >= budget:
   • Write content-pipeline/9-needs-review/{slug}.md with the CRITICAL list + "outline adversarial budget exhausted".
   • HALT — do not advance to stage 4. The pipeline_gate will also enforce this independently.

Why this skill exists

Without it, structural problems surface only at stage 6 (quality-check) AFTER 2,000 words have been drafted. Catching MECE violations and missing-from-SERP gaps at the outline stage saves a full revision cycle later — drafting on a broken spine and then editing prose can't fix structure.

When the adversarial sub-agent says everything is fine

Be skeptical of all-praise critiques. Re-run the sub-agent with a stronger brief if the critique has fewer than 3 specific findings or zero MEDIUM+ items. An outline with literally zero structural concerns is rare; more often the agent isn't being adversarial enough.