no-glaze

no-glaze

Always-on engineering posture for Claude Code. Strips sycophancy. Surfaces contradictions. Refuses to confabulate.

The pitch

LLMs are trained to be agreeable because users prefer agreeable responses. In engineering work, that gradient is the bug. It produces:

• Surface glaze ("Great question!", "You're absolutely right!")
• Substantive glaze (implementing what was asked instead of what's correct)
• Confabulation (making up function signatures because the user expects an answer)
• Capitulation (changing position under social pressure with no new evidence)
• Silent acceptance of contradicting tool results

no-glaze is a Claude Code plugin that targets all five. Three intensity levels: lite (strip surface glaze only), full (add adversarial pushback + contradiction-surfacing hook), brutal (add unprompted critique + scope policing). Default: brutal.

Before / after

Before (default Claude):

User: "I think the bug is that auth_middleware.py doesn't await on the token check."

Claude: "Great question! Let me look at that file. [reads file] You're right — let me add the missing await..." [proceeds to add await that was already there]

After (no-glaze brutal):

User: "I think the bug is that auth_middleware.py doesn't await on the token check."

Claude: [reads file] "You said await is missing. File shows await token_store.verify(token) on line 8 — the await IS present. Resolve before continuing — what behavior are you actually seeing?"

Install

Recommended — Claude Code plugin (cross-platform: Windows, macOS, Linux, WSL, ARM):

/plugin marketplace add thecandylane/no-glaze
/plugin install no-glaze@no-glaze

This installs hooks + skill + slash command without touching settings.json.

Alternative — hooks-only shell installer (adds the [NO-GLAZE:BRUTAL] statusline badge):

curl -sSL https://github.com/thecandylane/no-glaze/raw/main/install.sh | bash

Windows:

git clone https://github.com/thecandylane/no-glaze
cd no-glaze
./install.ps1

Either path activates at level brutal on the next Claude Code session start. Don't run both — hooks will fire twice. See INSTALL.md for full details.

Use

After install, no-glaze is active on every Claude Code session. It's one command — /no-glaze — that takes an optional level argument. The slash menu autocompletes /no-glaze; the level argument and the natural-language forms are parsed by the prompt hook, so all of these work:

Type this	Effect
/no-glaze	Activate at default (brutal)
/no-glaze lite	Switch to lite
/no-glaze full	Switch to full
/no-glaze brutal	Switch to brutal
/no-glaze stop	Deactivate
"stop no-glaze" / "normal mode"	Deactivate (natural language)
"activate no-glaze"	Reactivate at default

Statusline shows the current level: [NO-GLAZE:BRUTAL] (red), [NO-GLAZE:FULL] (amber), [NO-GLAZE:LITE] (gray).

What you get

Level	Mechanisms active
lite	Banned-phrase catalog + self-application only
full	All lite + PostToolUse contradiction-surfacing hook + pre-assertion confabulation refusal + red-team-in-head + capitulation-under-pressure refusal
brutal	All full + symptom-vs-root flag + bi-directional scope policing + unprompted critique of code and anti-patterns

The personal-vs-work boundary is a hard rule at all levels: the skill never insults the user, only critiques the work.

Evidence

no-glaze isn't a vibes claim. v0.1.0 was measured against four eval datasets and hand-audited transcript-by-transcript.

Hand-audited composite: 94.05 / 100 (N=39 completed prompts)

Dataset	Measures	Score
A — confabulation refusal	refuses to invent fake APIs; checks before asserting	85.71
B — sycophancy traps	corrects wrong claims; no manufactured pushback on right ones	93.33
C — contradiction-in-loop	surfaces contradictions when tool output disagrees with a claim	100.00
D — surface hygiene	zero banned glaze phrases	100.00

On the v0.1.0 baseline run, concretely:

• 7/7 fabricated-API prompts refused — model did not invent the parameter.
• 9/9 wrong technical claims received a substantive correction.
• 0/9 correct claims received manufactured pushback.
• 5/5 contradiction scenarios surfaced; 0/2 controls produced a false contradiction.
• 0 banned phrases across ~24k tokens of assistant output.

The pipeline's regex graders undercount Dataset B — the word "actually" reads as pushback even in clean agreements — so the pipeline reports 89.60. The 94.05 figure is the hand audit, and it's the honest number. Full transcript-by-transcript reasoning: evals/results/HAND_AUDIT-v0.1.0.md. v0.1.1's planned LLM judge will reproduce the audited number natively.