adversarial-agents

jm-skills

A Claude Code plugin marketplace hosting multiple independent plugins. Install the whole marketplace with one command, then pick which plugins to enable.

Install

Add the marketplace

/plugin marketplace add jasonm4130/claude-skills

Install individual plugins

Plugin	Description	Install command
adversarial-agents	Configurable adversarial panel review for any artefact	/plugin install adversarial-agents@jasonm4130-claude-skills
deep-research	Multi-source research via parallel sub-agents and synthesis	/plugin install deep-research@jasonm4130-claude-skills
session-retro	Interactive session retrospectives that capture learnings to memory	/plugin install session-retro@jasonm4130-claude-skills
handoff	Context-fill-triggered handoff skill that auto-loads on next session	/plugin install handoff@jasonm4130-claude-skills

For full details on each plugin, see its own README:

• plugins/adversarial-agents/README.md
• plugins/deep-research/README.md
• plugins/session-retro/README.md
• plugins/handoff/README.md

Note for handoff users: The handoff plugin requires an additional statusLine wiring step in your ~/.claude/settings.json — see plugins/handoff/README.md for the snippet.

Skills

adversarial-agents

Configurable adversarial panel review for any artefact — plans, code, design docs, prose, model outputs.

• Panel auto-selection by artefact type (plan → YAGNI/Premortem/Hidden Assumptions; code → Saboteur/New Hire/Security Auditor; prose → Clarity/Hostile Reader/Devil's Advocate; etc.)
• Pre-commitment gate — user defends the artefact before the panel attacks, neutralising leading-question sycophancy
• Shared adversary contract — mandatory ≥1 finding per persona, named anti-rationalization failure modes
• [CONVERGED] overlap promotion — critiques surfaced by 2+ personas walked first
• Verbatim-substance standing rule — parent quotes critic verbatim to resist capability-asymmetry drift
• Dog-with-bone walk with deadlock cap at 3 counter-pushes

Trigger phrases include grill me, stress-test this, red-team, adversarial review, panel critique, find holes.

deep-research

Multi-source research with DAG-planned dispatch, cost-aware model selection (Haiku critics + Sonnet synthesis + Opus orchestrator), and three-pass synthesis (critic → citation-quality judge → final-judge).

Follows the lead-researcher → parallel sub-agents → synthesis pattern from Anthropic's multi-agent research system, extended with 2025–26 patterns (plan-as-DAG, role-separated critic vs judge, asymmetric models).

Install

This plugin uses the Claude Code plugin marketplace mechanism. Add to your ~/.claude/settings.json:

{
  "extraKnownMarketplaces": {
    "jasonm4130-claude-skills": {
      "source": { "source": "github", "repo": "jasonm4130/claude-skills" }
    }
  },
  "enabledPlugins": {
    "jm-skills@jasonm4130-claude-skills": true
  }
}

Then run /plugins in Claude Code to sync. Skills will be invokable as jm-skills:adversarial-agents and jm-skills:deep-research, plus by their natural trigger phrases.

Repo layout

.claude-plugin/plugin.json       # plugin manifest
skills/
  adversarial-agents/
    SKILL.md                     # skill definition + frontmatter
    personas/
      yagni.md, premortem.md, hidden_assumptions.md   # plan panel
      saboteur.md, new_hire.md, security_auditor.md   # code panel
  deep-research/
    SKILL.md

License

MIT — see LICENSE.

Acknowledgements

adversarial-agents was prompted by Matt Pocock's grill-me skill. The panel-of-personas + severity-promotion pattern draws on Alireza Rezvani's adversarial-reviewer and zscole's adversarial-spec. The full research synthesis behind the design decisions lives in the originating dotfiles plan (docs/plans/2026-05-16-skills-overhaul-research.md).