sycophancy-stack

A Claude Code marketplace with two skills that fix the two most common ways AI coding agents fail you.

• deglaze — the retrospective fix. Makes Claude honestly audit its own claimed-complete work against 17 named under-delivery patterns instead of polishing a summary over half-finished tasks. Trigger: /deglaze, "did you do your best", "what did you skip", "stop glazing".
• deliberate — the forward-looking fix. Three-agent deliberation loop (Opinion-Drafter, Goal-Auditor, Impact-Maximizer) that converges on ONE recommendation instead of bouncing decisions back as a menu of options. Trigger: /deliberate, "you decide", "i aint picking", "converge for me".

Install (Claude Code)

Add the marketplace, then install one or both plugins:

/plugin marketplace add LuciferDono/sycophancy-stack
/plugin install deglaze@sycophancy-stack
/plugin install deliberate@sycophancy-stack

Why a stack?

The two skills cover the two failure modes independently and reinforce each other.

• Without deliberate: the model bounces every substantive decision back at you as a menu of options.
• Without deglaze: the model declares a task done with a polished summary that hides what it actually skipped.
• With both: decisions get made honestly, and completion claims get audited honestly. That's the whole stack.

Research

Both skills compose techniques from published multi-agent-reasoning and sycophancy-mitigation research. Citations in each repo's SKILL.md and README.

License

Marketplace itself is MIT. Each plugin ships under its own MIT license.

{
  "$schema": "https://anthropic.com/claude-code/marketplace.schema.json",
  "name": "sycophancy-stack",
  "description": "The two-skill sycophancy stack for Claude Code: deglaze catches when the model half-assed the task; deliberate stops it from bouncing decisions back as menus.",
  "owner": {
    "name": "LuciferDono"
  },
  "metadata": {
    "description": "Anti-sycophancy skills for Claude Code"
  },
  "plugins": [
    {
      "name": "deglaze",
      "source": {
        "source": "github",
        "repo": "LuciferDono/deglaze"
      },
      "description": "Strip the sycophancy. Makes Claude honestly audit its own claimed-complete work against 17 named under-delivery patterns instead of polishing a summary over half-finished tasks.",
      "version": "1.0.0",
      "author": {
        "name": "LuciferDono"
      },
      "homepage": "https://github.com/LuciferDono/deglaze",
      "repository": "https://github.com/LuciferDono/deglaze",
      "license": "MIT",
      "keywords": [
        "sycophancy",
        "self-critique",
        "chain-of-verification",
        "accountability",
        "prompt-engineering",
        "claude-skill",
        "anti-glaze"
      ],
      "category": "quality",
      "tags": [
        "skills",
        "sycophancy",
        "self-critique",
        "audit",
        "accountability"
      ],
      "strict": false
    },
    {
      "name": "deliberate",
      "source": {
        "source": "github",
        "repo": "LuciferDono/deliberate"
      },
      "description": "Converges on ONE call instead of a menu of options. Two modes: a three-agent sequential loop for value-tradeoff decisions, and a Persona Panel of 3-5 parallel independent experts for domain-expertise decisions.",
      "version": "1.1.0",
      "author": {
        "name": "LuciferDono"
      },
      "homepage": "https://github.com/LuciferDono/deliberate",
      "repository": "https://github.com/LuciferDono/deliberate",
      "license": "MIT",
      "keywords": [
        "deliberation",
        "multi-agent",
        "multi-agent-debate",
        "sycophancy",
        "decision-making",
        "prompt-engineering",
        "claude-skill"
      ],
      "category": "workflow",
      "tags": [
        "skills",
        "multi-agent",
        "decision-making",
        "sycophancy"
      ],
      "strict": false
    }
  ]
}