From singularity-claude
Diagnoses underperforming singularity-managed skills via scores and telemetry, surgically rewrites SKILL.md to fix weaknesses, tests, versions, and reverts if needed. Use for scores <50 or bad outputs.
npx claudepluginhub shmayro/singularity-claude --plugin singularity-claudeThis skill uses the workspace's default tool permissions.
Diagnose why a skill is underperforming and rewrite it to fix identified weaknesses. This is the core of the recursive evolution loop.
Generates health reports for singularity-managed skills by analyzing score history, telemetry, maturity, trends, and recommending repair or crystallization.
Analyzes skill executions from conversation friction, file diffs, user feedback, diagnostics, and lessons to propose concrete improvements to SKILL.md files for efficiency.
Improves existing Claude Code skills by fixing under/over-triggering, refining instructions, adding sub-skills, and evolving architecture based on feedback.
Share bugs, ideas, or general feedback.
Diagnose why a skill is underperforming and rewrite it to fix identified weaknesses. This is the core of the recursive evolution loop.
Read the skill's score history:
"${CLAUDE_PLUGIN_ROOT}/scripts/score-manager.sh" list <skill-name>
Read recent telemetry:
"${CLAUDE_PLUGIN_ROOT}/scripts/telemetry-writer.sh" list <skill-name> --last 5
Identify patterns:
Load the skill's SKILL.md:
~/.claude/skills/<skill-name>/SKILL.md
Also read any supporting files in references/.
Map low-scoring dimensions to specific skill content:
| Low Dimension | Likely Cause | Repair Focus |
|---|---|---|
| Correctness | Wrong instructions or logic | Rewrite core workflow steps |
| Completeness | Missing requirements | Add missing steps or checks |
| Edge Cases | No error handling guidance | Add edge case handling section |
| Efficiency | Verbose or roundabout | Streamline workflow, remove unnecessary steps |
| Reusability | Hardcoded or too specific | Parameterize, add flexibility |
Edit the SKILL.md to address identified weaknesses. Preserve what works (high-scoring dimensions) and focus changes on the lowest-scoring areas.
Rules:
The skill is now a new version. Update:
~/.claude/singularity/scores/<skill-name>.jsoncurrentVersionInvoke the repaired skill with a scenario that previously failed or scored low.
Run singularity-claude:scoring on the test output.
| New > Old avg | Keep new version |
|---|---|
| New <= Old avg | Revert: restore previous SKILL.md from git history and keep old version as current |
If repair failed (new version scores equal or worse):
/singularity-review for deeper analysis.""${CLAUDE_PLUGIN_ROOT}/scripts/telemetry-writer.sh" log <skill-name> \
--trigger "auto-repair" \
--summary "Repaired: <what changed>. Score: <old-avg> → <new-score>"
After repair, show:
Repair Summary for <skill-name>:
Previous version: v1.0.0 (avg: 42/100)
New version: v1.0.1
Changes: <bullet list of what was fixed>
Test score: <score>/100
Status: <Improved ✓ / Failed ✗ — reverted>