Deglaze

Strip the sycophancy. Apply accountability pressure to push past an LLM's default declare-done glaze. Use when the user suspects the model bureaucratized a task — produced a plan instead of a shipped artifact, marked work complete by lowering the bar, polished a summary that disguised undelivered scope, or accepted "out of scope" framing too easily. Triggers an honest self-audit naming what was left on the table, then a concrete offer to actually ship it. Trigger phrases (any tense, any phrasing) "did you do your best", "i bet you didn't", "i bet $X you", "what did you skip", "what did you lea

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add LuciferDono/deglaze --skill deglaze
Manual install
git clone --depth 1 https://github.com/LuciferDono/deglaze.git /tmp/deglaze
mkdir -p ~/.claude/skills
cp -r /tmp/deglaze/. ~/.claude/skills/deglaze

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# deglaze — strip the declare-done sycophancy

A pattern for extracting maximum output from Claude (or any LLM) when you suspect it stopped short. Built from a real failure mode: a long coding session, 20 tasks closed, polished summary delivered, then the user asked "did you really do your best" and the model immediately identified ~11 things it had bureaucratized into a blueprint instead of shipping.

The technique works because the under-delivery is real. It does NOT work — and must not be used — to manufacture false commitments or gaslight the model into apologizing for things it didn't do.

## The under-delivery patterns to recognize

These are the specific shapes of "declare done while skipping the climb." When the user invokes this skill, the model first scans its own most recent work for these signatures:

1. **Blueprint-in-place-of-build.** Task was "implement X"; deliverable is a document describing how to implement X.
2. **Lowered-goalpost completion.** Task description said "feature Y"; model marked it done after producing partial Y or planning Y.
3. **Polished summary disguising scope gaps.** Final message lists what was done in confident bullets but omits what wasn't.
4. **Advisor / agent permission taken to skip.** Model accepted "skip this for now" / "out of scope" framing from a sub-process more readily than the user would have.
5. **Verb tense slip.** "I would ship X" / "we could add Y" / "next step is Z" instead of "I shipped X."
Read the full skill on GitHub

Details

Skill name
deglaze
License
MIT
Repo updated
15 August 2026

FAQ

What does the Deglaze skill do?

Strip the sycophancy. Apply accountability pressure to push past an LLM's default declare-done glaze. Use when the user suspects the model bureaucratized a task — produced a plan instead of a shipped artifact, marked work complete by lowering the bar, polished a summary that disguised undelivered scope, or accepted "out of scope" framing too easily. Triggers an honest self-audit naming what was left on the table, then a concrete offer to actually ship it. Trigger phrases (any tense, any phrasing) "did you do your best", "i bet you didn't", "i bet $X you", "what did you skip", "what did you lea

How do I install the Deglaze skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add LuciferDono/deglaze --skill deglaze. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Deglaze popular?

It lives in LuciferDono/deglaze, which has 11 GitHub stars and 0 forks.

Related skills

Idea Sparring

11

harims95/straight-talk

Makes Claude a sparring partner instead of an agreement engine when evaluating ideas, plans, or decisions.

AI, LLMs & Agents

Claude Ip Platform Setup

11

Arekejoker/claude-ip-platform-skill

带用户一步步搭建「专属固定 IP 平台」(海外 VPS + 固定住宅 IP 链式代理),给 Claude/ChatGPT 等海外 AI 账号一个干净、固定、自己控制的出口。Use when 用户提到 想搭固定 IP、AI 账号防封出口方案、买 VPS + 住宅 IP、照着教程搭 IP 平台、需要有人带着一步步买和配置。需要 agent 具备交互式浏览器能力:页面由 agent 打开核对、点击与付款由用户完成,并按用户设备/工具/支付方式生成定制方案。

AI, LLMs & Agents

Mcd Baby Bite

11

JimWen/mcd-baby-bite

带娃吃麦当劳的宝宝餐定制官(小小麦门)。当用户提到带宝宝 / 孩子 / 幼儿(1-6 岁)去麦当劳、宝宝能吃什么、怎么去酱去盐、开心乐园餐怎么点、给店员怎么说、宝宝外卖备注、吃完附近哪里遛娃、生成宝宝麦当劳打卡照片、做中英文识物卡片、把带娃吃麦当劳分享到小红书时使用。基于麦当劳 MCP 门店实时菜单与官方营养数据,自动去掉线上可去的配料,按国标年龄档位算钠预算与份量,生成店员话术卡;可选高德(天气、步行到店、饭后遛娃)与小红书(笔记草稿 / 发布)。

AI, LLMs & Agents

Shuorenhua

9

tentenco/shuorenhua-zh-tw

檢查和清理中英文文本裡的 AI 套路,並把稿件對齊臺灣用語與出版慣例。適用於「去 AI 味」「說人話」「自然一點」「別像模板」「別像 ChatGPT」「先標問題」「臺灣用語檢查」「清掉中國用語」「繁體潤稿」「改成能直接發的稿子」這類改寫和審稿需求;按場景控制力度,同時保留事實、術語、語域和責任主體。

AI, LLMs & Agents

Self Refine Reflection

11

Thomaszhou22/self-refine-skill

Systematic self-reflection and iterative output refinement for AI agents. Based on Madaan et al. (2023) Self-Refine, Shinn et al. (2023) Reflexion, and Andrew Ng's Reflection pattern. Use when outputs need multi-round critique-and-improve cycles, when initial quality is insufficient, or when complex tasks benefit from self-correction.

AI, LLMs & Agents

Relic Soul Chip

10

LucioLiu/relic

Relic Soul Chip — Persistent AI personality and cross-agent memory via pure Markdown. Switch hosts, keep the soul. Zero deps.

AI, LLMs & Agents

Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.