Claude Skill

insight-adversarial

Multi-pass adversarial critique from dynamically selected personas — including the current reviewer's own view. Trigger on: "adversarial", "red team this", "multi-POV review", "stress-test", "poke holes", "what am I missing", "attack my plan", "critique from multiple angles", "is

LLM Mart · 0 points · 0 views 0 listing impressions 0 install-command copies
Download allemaar-open-skills-skills_insight-adversarial-acf5f79.zip · 9 KB

Install

skills CLI npx skills add https://github.com/allemaar/open-skills/tree/main/skills/insight-adversarial
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install allemaar-open-skills@llmmart
Git git clone https://github.com/allemaar/open-skills.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole allemaar/open-skills collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

/insight-adversarial

Structured adversarial review that assembles a panel of distinct critic personas, runs each through a focused pass, and synthesizes a verdict. Personas are selected based on the nature of the target — a YON workflow gets different critics than a business plan or a UX design.

Structured execution spec: protocol.yon. Read it for the canonical rules and step sequence; this file is explanation. The two must stay in sync — if you edit one, update the other and refresh the @STAMP date.

Caller Options. Before executing, run the Caller Options protocol (caller-options/SKILL.md): triage this invocation for material optionality across the venues and modes declared in front-matter; if one path clearly dominates, proceed silently; otherwise surface the options to the caller. A resolved-invocation marker means COP already ran — execute the fixed combination directly, do not re-enter COP.

How Passes Work

Pass What Happens
Pass 1 Classify the target → select 4–5 personas → run each → synthesize verdict
Pass 2 Check against Pass 1 findings → add 3–4 new personas not used before → re-run → delta verdict
Pass 3 Final wrap-up. If all critical issues resolved or within params, declare done. Otherwise escalate.

State tracking: at the top of each pass, list which personas were used in prior passes. New passes must not reuse substantive critic personas; Current Reviewer's Own View recurs in every pass.

Step 1 — Target Classification

Before selecting personas, classify the target:

Target Type Key Question Primary Risk
plan Is this executable? Scope drift, undeclared deps
design / architecture Does structure match intent? Coupling, ownership violations
yon-workflow Does this execute correctly end-to-end? Gate logic, step ordering, missing catches
document / spec Is this complete and unambiguous? Gaps, contradictions, assumed context
idea / concept Is the core premise sound? Wrong problem, premature solution
business / product Is this viable and desirable? Market assumptions, cost blindness
code / implementation Is this correct and maintainable? Edge cases, coupling, test coverage
ui / ux Will real users succeed with this? Flow breakdowns, mental model gaps

Load personas/roster.md to see all available personas and their best-used-for tags. Use the classification to select the most relevant 4–5 personas for this pass.

Step 2 — Persona Selection Logic

  1. Pick the primary persona — whose domain most directly overlaps the target's failure mode
  2. Pick 2–3 supporting personas — different angles that complement, not duplicate, primary
  3. Always include Current Reviewer's Own View as the final persona in every pass
  4. Never reuse a substantive critic persona across passes; Current Reviewer's Own View is exempt — track used personas explicitly

Step 3 — Run Each Persona Pass

For each selected persona:

  1. State the persona name and their lens in one sentence
  2. Report specific, evidenced findings. If none are supported, say no material finding supported in the inspected scope, naming that scope and remaining uncertainty. Never invent an issue to fill the format.
  3. Merge convergent findings across personas, retaining their evidence; do not invent differences to make each lens unique.

Format:

## 🔴 [PERSONA NAME]
*Lens: [one-sentence description]*

[Critique — specific, no softening]

Current Reviewer's Own View always goes last. First person, clear position, no hedging.

Step 4 — Synthesis

  • Critical Issues — up to 3 supported issues, ranked by actual severity; label lower severities honestly and say when no critical issue is supported. A sound target may have no material finding.
  • Blind Spots — what the aggregate reveals that no single persona fully caught
  • Remediation — minimum surgical fix per Critical Issue. If unfixable, say so.
  • One-Line Verdict — honest, specific, no hedging

Step 5 — Pass Wrap Logic

  • Pass 1 or 2: Ask user if they want another pass. If yes → load new personas, run delta critique focused on unresolved issues.
  • Pass 3: Check if all Critical Issues resolved or explicitly accepted. If yes → final consolidated verdict + declare wrap. If no → flag remaining open issues, recommend escalation (e.g., /plan-create).

Rules

  • MUST classify target type before selecting personas
  • MUST select minimum 4 personas per pass (including Current Reviewer's Own View)
  • MUST include Current Reviewer's Own View in every pass, always last
  • MUST NOT reuse a substantive critic persona across passes; Current Reviewer's Own View is exempt
  • MUST NOT offer solutions inside a persona pass — remediation belongs in Step 4 only
  • MUST NOT soften findings to be diplomatic
  • SHOULD add 3–4 genuinely new personas on Pass 2
  • SHOULD tailor persona selection to target type
  • MAY suppress a persona if user explicitly requests it

Files

Human output. This skill's handler-facing output obeys the human-output contract (human-output/SKILL.md).

Next skills. On completion, run the Next Skills protocol (next-skills/SKILL.md): surface the next-skills recommendations from front-matter for the caller to pick. Offer only — never auto-invoke.

Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): apply its signal taxonomy and threshold to this run; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.

Files (open-skills)
  • personas
    • roster.md 8.7 KB
      Loading…
  • protocol.yon 8.3 KB · in bundle
  • SKILL.md 7.4 KB
    ---
    name: insight-adversarial
    description: >
      Multi-pass adversarial critique from dynamically selected personas — including the current reviewer's own view.
      Trigger on: "adversarial", "red team this", "multi-POV review", "stress-test", "poke holes",
      "what am I missing", "attack my plan", "critique from multiple angles", "is this solid",
      "what would a skeptic say", or any request for rigorous structured criticism of a plan, design,
      doc, workflow, YON, idea, or system. Runs 1–3 passes. First pass selects 4+ personas from the
      roster. Each subsequent pass adds 3–4 new personas to the mix. Third pass wraps with a final
      verdict if all issues are within acceptable params. Do not skip this skill when the user wants
      real scrutiny — "just give me feedback" often means they need this.
    caller-options:
      venue: [inline, delegated]
      default-policy: ask
    visibility: public
    self-improvable: true
    next-skills:
      - skill: plan-create
        phrase: "/plan-create"
        why: "Redesign the plan if the stress-test found fundamental flaws"
      - skill: plan-execute
        phrase: "/plan-execute"
        why: "Proceed to execution if the plan held up under scrutiny"
      - skill: human-rewrite
        phrase: "/human-rewrite"
        why: "Multi-pass persona output arrives long and in several voices — condense it to the objections that survived, keeping each one's force intact"
    triggers:
      - "/insight-adversarial"
      - "adversarial"
      - "red team this"
      - "multi-POV review"
      - "stress-test"
      - "poke holes"
      - "what am I missing"
      - "attack my plan"
      - "critique from multiple angles"
      - "is this solid"
      - "what would a skeptic say"
    ---
    
    # /insight-adversarial
    
    Structured adversarial review that assembles a panel of distinct critic personas, runs each through a focused pass, and synthesizes a verdict. Personas are selected based on the *nature of the target* — a YON workflow gets different critics than a business plan or a UX design.
    
    > **Structured execution spec:** [`protocol.yon`](protocol.yon). Read it for the canonical rules and step sequence; this file is explanation. The two must stay in sync — if you edit one, update the other and refresh the `@STAMP` date.
    
    > **Caller Options.** Before executing, run the Caller Options protocol (`caller-options/SKILL.md`): triage this invocation for material optionality across the venues and modes declared in front-matter; if one path clearly dominates, proceed silently; otherwise surface the options to the caller. A resolved-invocation marker means COP already ran — execute the fixed combination directly, do not re-enter COP.
    
    ## How Passes Work
    
    | Pass | What Happens |
    |------|-------------|
    | **Pass 1** | Classify the target → select 4–5 personas → run each → synthesize verdict |
    | **Pass 2** | Check against Pass 1 findings → add 3–4 *new* personas not used before → re-run → delta verdict |
    | **Pass 3** | Final wrap-up. If all critical issues resolved or within params, declare done. Otherwise escalate. |
    
    **State tracking:** at the top of each pass, list which personas were used in prior passes. New passes must not reuse substantive critic personas; Current Reviewer's Own View recurs in every pass.
    
    ## Step 1 — Target Classification
    
    Before selecting personas, classify the target:
    
    | Target Type | Key Question | Primary Risk |
    |-------------|-------------|--------------|
    | `plan` | Is this executable? | Scope drift, undeclared deps |
    | `design / architecture` | Does structure match intent? | Coupling, ownership violations |
    | `yon-workflow` | Does this execute correctly end-to-end? | Gate logic, step ordering, missing catches |
    | `document / spec` | Is this complete and unambiguous? | Gaps, contradictions, assumed context |
    | `idea / concept` | Is the core premise sound? | Wrong problem, premature solution |
    | `business / product` | Is this viable and desirable? | Market assumptions, cost blindness |
    | `code / implementation` | Is this correct and maintainable? | Edge cases, coupling, test coverage |
    | `ui / ux` | Will real users succeed with this? | Flow breakdowns, mental model gaps |
    
    Load `personas/roster.md` to see all available personas and their best-used-for tags. Use the classification to select the most relevant 4–5 personas for this pass.
    
    ## Step 2 — Persona Selection Logic
    
    1. Pick the **primary persona** — whose domain most directly overlaps the target's failure mode
    2. Pick 2–3 **supporting personas** — different angles that complement, not duplicate, primary
    3. Always include **Current Reviewer's Own View** as the final persona in every pass
    4. **Never reuse** a substantive critic persona across passes; Current Reviewer's Own View is exempt — track used personas explicitly
    
    ## Step 3 — Run Each Persona Pass
    
    For each selected persona:
    
    1. State the persona name and their lens in one sentence
    2. Report specific, evidenced findings. If none are supported, say **no material finding supported in the inspected scope**, naming that scope and remaining uncertainty. Never invent an issue to fill the format.
    3. Merge convergent findings across personas, retaining their evidence; do not invent differences to make each lens unique.
    
    **Format:**
    
    ```
    ## 🔴 [PERSONA NAME]
    *Lens: [one-sentence description]*
    
    [Critique — specific, no softening]
    ```
    
    Current Reviewer's Own View always goes last. First person, clear position, no hedging.
    
    ## Step 4 — Synthesis
    
    - **Critical Issues** — up to 3 supported issues, ranked by actual severity; label lower severities honestly and say when no critical issue is supported. A sound target may have no material finding.
    - **Blind Spots** — what the aggregate reveals that no single persona fully caught
    - **Remediation** — minimum surgical fix per Critical Issue. If unfixable, say so.
    - **One-Line Verdict** — honest, specific, no hedging
    
    ## Step 5 — Pass Wrap Logic
    
    - **Pass 1 or 2:** Ask user if they want another pass. If yes → load new personas, run delta critique focused on unresolved issues.
    - **Pass 3:** Check if all Critical Issues resolved or explicitly accepted. If yes → final consolidated verdict + declare wrap. If no → flag remaining open issues, recommend escalation (e.g., `/plan-create`).
    
    ## Rules
    
    - MUST classify target type before selecting personas
    - MUST select minimum 4 personas per pass (including Current Reviewer's Own View)
    - MUST include Current Reviewer's Own View in every pass, always last
    - MUST NOT reuse a substantive critic persona across passes; Current Reviewer's Own View is exempt
    - MUST NOT offer solutions inside a persona pass — remediation belongs in Step 4 only
    - MUST NOT soften findings to be diplomatic
    - SHOULD add 3–4 genuinely new personas on Pass 2
    - SHOULD tailor persona selection to target type
    - MAY suppress a persona if user explicitly requests it
    
    ## Files
    
    - [personas/roster.md](personas/roster.md) — Full persona library. Load during Step 1 to inform selection.
    
    > **Human output.** This skill's handler-facing output obeys the human-output
    > contract (`human-output/SKILL.md`).
    
    > **Next skills.** On completion, run the Next Skills protocol (`next-skills/SKILL.md`): surface the `next-skills` recommendations from front-matter for the caller to pick. Offer only — never auto-invoke.
    
    > **Self-improvement.** On completion, run the Self-Improvement Protocol (`self-improve/SKILL.md`): apply its signal taxonomy and threshold to this run; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.