qult
Quality by Structure, Not by Promise. A harness engineering tool that enforces code quality through walls, not words.
Prompts are suggestions. Hooks are enforcement. qult uses 7 hooks + MCP server + 4-stage independent review to block quality regressions with exit 2 (DENY), not advisory messages. Distributed as a Claude Code Plugin. Install with
/plugin install.
Why harness engineering?
AI coding agents are powerful but unreliable at self-regulation.
- They leave lint errors behind and move to the next file.
- They commit without running tests.
- They praise their own code and call the review done.
Anthropic's Harness Design article identifies two root causes: context coherence loss (models degrade as context fills) and self-evaluation bias (agents praise their own mediocre work). The solution: separate generation from evaluation, and enforce quality structurally.
The broader discipline of harness engineering, formalized by Martin Fowler, frames this as:
Agent = Model + Harness. Enforce invariants, don't micromanage implementations.
Research validates this approach:
- TDAD (arXiv:2603.17973): Adding TDD instructions via prompts alone increased regressions from 6% to 10%. Structural enforcement reduced them to 1.8%.
- Specification as Quality Gate (arXiv:2603.25773): AI reviewing AI is structurally circular. Correlated errors echo rather than cancel. Deterministic gates must come first, AI review only for the residual.
qult implements this architecture:
Deterministic gates (lint, typecheck) → Executable specifications (test) → AI review (residual only)
Trust structure, not models
| Approach | Mechanism | Guarantee |
|---|---|---|
| Prompt-based (superpowers, etc.) | Persuasive skill prose | None. Model may skip steps |
| Structure-based (qult) | Hook exit 2 (DENY) | Structural. Cannot be bypassed by the model |
In Fowler's taxonomy: qult's hooks are sensors (feedback controls that observe and correct), while skills and CLAUDE.md are guides (feedforward controls that steer before action). Both are necessary; qult provides both.
Philosophy
Quality by Structure, Not by Promise.
1. The Wall doesn't negotiate.
Prompts are suggestions. Hooks are enforcement. Quality is not left to promises.
2. The architect designs, the agent implements.
Humans decide what to build. AI decides how to build it.
Ambiguity is resolved by asking, never by guessing.
3. Proof or Block.
"Done" is not evidence. Tests pass, review passes — then it's done.
Claims without evidence are structurally blocked.
4. fail-open — cult of quality, humble about ourselves.
qult's own failures never block Claude. Break? Open the gate.
Fanatical about quality. Humble about our tools.
[!NOTE] You may see
SessionStart:startup hook errororStop hook errorat session start. This is not a qult bug. It's a known Claude Code UI bug that misreports hook success/failure (#12671, #21643, #10463). Hooks are working correctly.
[!WARNING] PreToolUse DENY may be ignored. qult correctly returns
exit 2, but Claude Code sometimes executes the tool anyway (#21988, #4669, #24327). Waiting for a Claude Code fix.
Japanese README / README.ja.md
How it works
flowchart LR
Edit["Edit / Write"] --> Gate{"Gate\n(lint, type)"}
Gate -- pass --> OK["Continue"]
Gate -- fail --> PF["pending-fixes"]
PF --> Next["Try to Edit\nanother file"]
Next --> DENY["The Wall\n(DENY exit 2)"]
DENY --> Fix["Fix the\nsame file"]
Fix --> Gate
style DENY fill:#e67e80,color:#2d353b,stroke:#e67e80
style OK fill:#a7c080,color:#2d353b,stroke:#a7c080
style PF fill:#dbbc7f,color:#2d353b,stroke:#dbbc7f
Operates on the Generator-Evaluator pattern from Anthropic's Harness Design article: