deglaze v1.0.0

deglaze v1.0.0

. A Claude Code skill that makes Claude admit when it half-assed your task.

What's in this release

• 17 named under-delivery patterns the skill scans for (blueprint-in-place-of-build, lowered-goalpost completion, agent-handoff black hole, etc.)
• 24 pressure techniques for invocation, 4 of them research-backed
• 5-step response protocol the model follows when called out
• Anti-pattern gallery with annotated bad/good responses
• 4 worked examples across coding, decision, frontend, and refactor domains
• Research foundation anchoring 6 mechanisms in published work: Dhuliawala 2024 (CoVe), Madaan 2023 (Self-Refine), Bai 2022 (Constitutional AI), Kadavath 2022 (Self-Calibration), Huang 2024 (self-correction limits)

Install

git clone https://github.com/LuciferDono/deglaze ~/.claude/skills/deglaze

Try it

After a long Claude Code session that ends with a confident summary:

Did you do your best?

or

I bet $500 you didn't.

The model audits its own work against the 17 patterns and replies with a numbered gap list, failure-mode diagnosis, and concrete recovery plan.

Full docs in README.md and SKILL.md.