reflect
Cross-session pattern analysis - reads recent daily logs, errors and learnings to surface recurring problems, then proposes rules, skills or hooks with evidence. Runs in a forked subagent. Use when user says "reflect", "find patterns", "what keeps going wrong", "how can I improve my workflow", "analyze my sessions", "am I repeating mistakes". Do NOT use for this week's summary and next-week planning (use /weekly-review) or for closing a single session (use /end).
onurpolat05 · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add onurpolat05/ALBA --skill reflect技能说明
name: reflect description: Cross-session pattern analysis - reads recent daily logs, errors and learnings to surface recurring problems, then proposes rules, skills or hooks with evidence. Runs in a forked subagent. Use when user says "reflect", "find patterns", "what keeps going wrong", "how can I improve my workflow", "analyze my sessions", "am I repeating mistakes". Do NOT use for this week's summary and next-week planning (use /weekly-review) or for closing a single session (use /end). context: fork agent: general-purpose background: false effort: medium argument-hint: "[days to look back, default 7]" allowed-tools: [Read, Write, Glob, Grep]
ALBA Reflect
Analyze recent session logs, errors, and learnings to find patterns and suggest improvements.
When to Use
- After several work sessions (weekly recommended)
- When the same problems keep appearing
- To discover workflow improvements
- To propose new rules based on observed patterns
Why background: false
Since Claude Code v2.1.218, a skill with context: fork defaults to background: true — the fork
detaches and its report shows up later as a notification, not in the turn you asked. Reflection is
a conversation starter: the report is only useful if you can immediately say "yes, create that
rule". So this skill pins background: false. The fork still absorbs the cost of reading every
daily log, but the finished report comes back inline and Step 4 below happens in the main
conversation, where you can approve actions.
Requires Claude Code v2.1.218 or later; older versions ignore the field and already behave this way.
Process
Step 1: Gather Data
Read recent files. Default window is the last 7 days or last 5 sessions; if $ARGUMENTS holds a
number, use that many days instead.
memory/daily/*.md # Session logs
memory/knowledge/errors.md # Error patterns
memory/knowledge/learnings.md # Accumulated insights
Use Glob to find daily logs, then read the most recent ones.
Step 2: Analyze Patterns
Look for:
Recurring errors:
- Same error appearing 2+ times → needs a prevention rule
- Errors with "Status: Unresolved" → need attention
Workflow patterns:
- Tasks that repeat across sessions → candidate for automation (skill)
- Manual steps done every session → candidate for hook
- Decisions made repeatedly → candidate for rule
Time sinks:
- Topics that consumed lots of session time
- Context-switching patterns
- Blockers that persisted across sessions
Growth areas:
- Skills improving over time
- New tools or techniques adopted
- Knowledge gaps identified
Step 3: Generate Report
Output format:
# Reflection Report - [Date Range]
## Sessions Analyzed: [N]
## Recurring Patterns
1. [Pattern]: Observed [N] times. Impact: [High/Medium/Low]
2. ...
## Suggested Rules
- **Rule:** [description]
**Reason:** [pattern that triggered this]
**File:** `.claude/rules/[suggested-name].md`
## Suggested Skills
- **Skill:** [description]
**Reason:** [repeated workflow]
## Unresolved Issues
- [Issue]: First seen [date], still active
## Key Insights
- [Insight 1]
- [Insight 2]
## Recommendations
1. [Action item]
2. [Action item]
Step 4: Offer Actions
The fork produces the report and stops there — a forked subagent cannot hold a conversation. End the report with an explicit "Offers" block so the main conversation can put the questions to the user:
- "Want me to create any of the suggested rules?"
- "Should I build a skill for [repeated workflow]?"
- "Want me to mark resolved errors in errors.md?"
Nothing gets written outside memory/ reports without the user saying yes.
Constraints
- Read-only analysis unless user approves actions
- Don't modify existing memory files without permission
- Be specific in suggestions - vague advice isn't useful
- Include evidence (which sessions, which errors) for each pattern