Corpus maps claude-reflect to automatic correction capture with confidence scoring and human review before CLAUDE.md writes.
Tool profile · provisional profile
claude-reflect
A useful correction-capture pattern. It starts at the right place, normal work, but needs promotion, authority, and sharing policy around it.
Provisional fit
63/100
Best for: Developers who want repeated corrections to become reviewable agent instructions instead of disappearing after a session.
Avoid if: you need a fully governed, citation-complete knowledge architecture without adding policy, evidence capture, and review workflow around the tool.
Caution: Capturing corrections is not enough. Bad or one-off corrections need review thresholds, provenance, and expiry.
Model signature: Production primary · developer scope · Hook-based capture
Layer coverage
Where this tool fits.
This is not a completed review. It is a provisional profile from public positioning plus known failure-mode mapping. Hands-on benchmarks, source snapshots, and citation-bound claims are still required before stronger conclusions.
Production
Curation
Storage
Activation
Governance
Review rule: a tool does not get credit for a layer unless it exposes inspectable behavior, not just a marketing claim.
Evidence notes
What the provisional profile has applied so far.
The latest model treats this as production plus curation: capture candidates, then decide blast radius and promotion path.
Needs review for false positives, prompt-injection resistance, and how often captured rules actually improve future runs.
Review packet
What a complete review must contain.
This page exposes the intended review structure. The current artifact is a profile, not a completed evidence-backed review.
Canonical source
Strengths
Developers who want repeated corrections to become reviewable agent instructions instead of disappearing after a session.
Limitations
Capturing corrections is not enough. Bad or one-off corrections need review thresholds, provenance, and expiry.
Dimension assessment
Scope, volatility, authority, lifecycle, resource economics, interoperability, and evidence quality must each get a rationale and citations before final scoring.
Open questions
- What can be verified from docs, code, issues, benchmarks, and changelogs?
- Where does the tool fail under stale, contradictory, private, or high-cost knowledge?
- Which claims are vendor claims versus independently observed behavior?
Benchmark critique
No benchmark number is accepted as architectural evidence unless it says which layer it tests and what it misses: lifecycle, scope boundaries, authority, context cost, and governance.
Related systems
Related tools should be connected by evidence-backed edges: competes with, integrates with, implements concept, evaluated by, or has governance gap.
Update history
Provisional profile created. Stale-review detection, source snapshots, and changelog watching are required before this becomes a durable review.