"LLMs often 'satisfice' (accept good-enough) rather than optimize." — Claude
A Claude Code plugin that makes Claude evaluate its own work quality before declaring any task complete. Instead of accepting the first working solution, Claude scores its output against a 7-dimension rubric and improves anything that falls short.
This skill was itself self-scored and improved after it was created — the initial implementation scored below the 9.5 threshold, triggered the improvement loop, and was refined until it passed. The tool practices what it preaches.
Without a quality gate, Claude tends to stop at "it works" rather than "it's good." Self-score adds a structured reflection step that catches:
- Code that works but is messy or over-engineered
- Solutions that meet the literal request but miss the user's actual intent
- Missing edge cases, security gaps, or performance issues
- Incomplete implementations that skip implicit requirements
Every task is scored across 7 dimensions (1.0–10.0 each):
| Dimension | What It Measures | 10 (Excellent) | 5 (Mediocre) |
|---|---|---|---|
| Correctness | Works as intended, edge cases handled | Bulletproof, no logical errors | Partially works, known issues |
| Code Quality | Clean, readable, idiomatic | Senior engineer approves without comments | Functional but messy |
| Completeness | All requirements addressed | Every requirement met, including implicit ones | Significant gaps |
| Elegance | Right abstraction, not over-engineered | Nothing to add, nothing to remove | Over- or under-abstracted |
| Performance | Efficient, no wasted resources | Optimal for the context | Noticeable issues |
| Security | No vulnerabilities, inputs validated | No attack surface added | Potential vulnerabilities |
| User Intent Alignment | Solves what was meant, not just literal | Spirit and letter nailed | Technically correct, misses the point |
- Score — After completing work, Claude reviews each dimension and assigns a score
- Evaluate — If the overall average is >= 9.5, the task is complete
- Improve — If below 9.5, Claude identifies specific targeted fixes for each weak dimension
- Iterate — With user approval, improvements are implemented and re-scored
The plugin also detects red flags like score inflation (every dimension 9.5+), N/A abuse, and lazy identical scores across dimensions.
Clone this repo into your Claude Code plugins directory:
git clone https://github.com/augmentedivan/claude-self-score.git ~/.claude/plugins/self-scoreThat's it. The skill activates automatically when Claude is about to declare a task complete.
For stronger enforcement, add this line to your ~/.claude/CLAUDE.md (or project-level CLAUDE.md):
- **Self-score on completion** — invoke the `self-score` skill before declaring any task complete. Skip for non-task interactions (questions, explanations, lookups).This makes the self-score step an explicit rule rather than relying solely on the skill's auto-trigger.
After completing a task, Claude produces a scorecard like this:
## Self-Score
| Dimension | Score |
|------------------------|-------|
| Correctness | 9.5 |
| Code Quality | 8.5 |
| Completeness | 9.5 |
| Elegance | 9.0 |
| Performance | N/A |
| Security | N/A |
| User Intent Alignment | 9.5 |
| **Overall** | **9.2** |
### Improvements to reach 9.5+
| Dimension | Current | Target | What to change |
|-----------|---------|--------|----------------|
| Code Quality | 8.5 | 9.5 | Extract repeated validation logic into a shared helper |
| Elegance | 9.0 | 9.5 | Replace nested conditionals with early returns |
Want me to implement these improvements?
- Claude Code CLI
MIT
Created by Ivan