review
by alirezarezvanireview audits Claude Code auto-memory in MEMORY.md to find promotion candidates, stale references, duplicates, consolidation opportunities, and health checks for Context Engineering workflows.
This skill scores 78/100, making it a solid listing candidate for directory users who use Claude Code auto-memory and want an agent-guided audit workflow. It provides clear triggers, command variants, and concrete analysis steps, though it is somewhat narrow and relies on assumed local memory paths rather than packaged tooling.
- Clear triggerability: the description explicitly says to use it for `/si:review` or when users ask what has been learned and what should be promoted or pruned.
- Operational workflow is concrete: it includes usage modes such as `--quick`, `--stale`, and `--candidates`, plus step-by-step guidance for locating and analyzing `MEMORY.md`.
- Provides agent leverage beyond a generic prompt by defining recurrence, staleness, consolidation, and health checks for auto-memory review.
- Narrow fit: it is mainly useful for repositories using the expected Claude Code auto-memory layout under `~/.claude/projects/.../memory`.
- No support scripts, references, README, or install command are present, so users must rely entirely on the SKILL.md instructions.
Overview of review skill
What review does
review is a self-improving-agent skill for auditing Claude Code auto-memory. It helps an agent inspect MEMORY.md, find promotion candidates, identify stale or duplicated entries, and produce maintenance recommendations instead of letting project memory grow unchecked. Use the review skill when you want to know what the agent has learned, what should become durable project guidance, and what should be pruned.
Best fit for this review skill
This skill is most useful for teams using Claude Code memory as part of a Context Engineering workflow. It fits projects where repeated lessons, tool preferences, architectural decisions, or recurring mistakes are captured automatically but need periodic editorial review. It is less useful if your project does not use Claude Code memory, has no .claude project memory directory, or only needs a one-off summary of a source file.
What makes it different from a generic prompt
A generic “summarize my memory” prompt usually misses operational checks. The review skill gives the agent a concrete audit path: locate the project-specific memory directory, read MEMORY.md, check line-count pressure against the 200-line startup limit, detect recurrence, verify stale file references, and separate promotion candidates from low-value entries. That structure makes the output more actionable for maintaining long-running agent context.
How to Use review skill
review install and first files to inspect
Install the skill from the GitHub repository with:
npx skills add alirezarezvani/claude-skills --skill review
After install, read SKILL.md first. This repository path contains the complete workflow in a single file, with no extra rules/, resources/, references/, or scripts to inspect. The most important source sections are Usage, What It Does, Step 1: Locate memory directory, and Step 2: Read and analyze MEMORY.md.
Basic review usage commands
The upstream skill is designed around /si:review style invocation:
/si:reviewfor a full memory audit/si:review --quickfor counts and top candidates/si:review --staleto focus on outdated entries/si:review --candidatesto show only likely promotions
In practice, call the review skill from the project root you want audited. The skill expects access to Claude Code’s project memory, usually under a path derived from the current working directory inside ~/.claude/projects/.../memory.
Inputs that improve review output
A weak prompt is: “Review memory.”
A stronger prompt is:
“Run the review skill for this repository. Check the project auto-memory, count MEMORY.md lines against the startup limit, identify recurring lessons that should be promoted into durable instructions, flag stale entries that reference deleted files or old tools, and group recommendations by promote, consolidate, prune, and keep.”
This works better because it names the audit criteria and the decision buckets. If you know recent project changes, include them: renamed directories, removed tools, new coding standards, or policy changes. Those details help the agent judge whether a memory entry is stale or still valuable.
Suggested review workflow
Start with /si:review --quick if the memory file is large or you only need triage. Run the full review when preparing a context refresh, onboarding a new agent setup, or cleaning accumulated memory before a major feature cycle. Use --stale after large refactors, dependency changes, or file moves. Use --candidates when your goal is Context Engineering: promoting repeated lessons into CLAUDE.md, project docs, or team rules.
review skill FAQ
Is review only for Claude Code memory?
Yes, the skill is specifically shaped around Claude Code auto-memory conventions, especially project memory under ~/.claude/projects/.../memory and MEMORY.md. You can adapt the review guide to other memory files, but its strongest value comes when the agent can follow the expected Claude Code memory layout.
What happens if no memory directory exists?
The skill should report that auto-memory may be disabled or unavailable, then suggest checking memory setup, such as using /memory. This is a useful pre-install consideration: if your team has not enabled or accumulated Claude Code memory, review will have little to analyze until memory exists.
When should I not use review?
Do not use review as a replacement for code review, security review, or documentation QA. It audits agent memory quality, not source code correctness. It is also a poor fit for brand-new projects with no recurring patterns, teams that deliberately avoid persistent agent memory, or cases where you only need a simple summary of a single document.
How does review for Context Engineering help?
For Context Engineering, review turns raw memory into a maintenance queue. It helps decide which repeated observations should become stable context, which entries are duplicates, and which outdated notes should be removed before they mislead future agents. The output is most valuable when you treat it as an editorial pass over your agent’s operating context.
How to Improve review skill
Give review evidence, not just a command
The review skill performs better when the agent can inspect the actual memory files and the current repository. Make sure it can verify referenced paths with filesystem checks rather than relying only on memory text. If you suspect stale guidance, say why: “we migrated from Jest to Vitest,” “the api/legacy folder was removed,” or “the deployment process changed last week.”
Ask for decision-ready categories
Request categories that map to action. Good buckets include:
- Promote to durable project instruction
- Consolidate with similar entries
- Prune because stale or low-value
- Keep as temporary memory
- Investigate because evidence is unclear
This prevents the review output from becoming a vague summary and makes it easier to update CLAUDE.md, team docs, or memory files after the audit.
Common failure modes to watch for
The most common failure is over-promoting one-off observations. A memory item should usually be promoted only if it is recurring, policy-like, or broadly useful across future sessions. Another failure is pruning entries just because they are old; age alone is not staleness. Better stale signals include missing files, replaced tools, contradicted conventions, or references to completed temporary work.
Iterate after the first review
After the first review, ask the agent to convert recommendations into a minimal patch plan: what to add, merge, remove, and re-check. Then run review again in quick mode to verify that the memory is shorter, less duplicated, and better aligned with current project behavior. This makes review an ongoing maintenance loop rather than a one-time report.