ice-review

ICE review workflow for consolidating knowledge from completed tasks into reusable memory and future leverage.

SKILL.md

GitHub

ICE Review Process

Description

A cross-task self-evolution skill based on the ICE (Investigate-Consolidate-Exploit) strategy from arXiv:2401.13996. This skill enables systematic knowledge extraction and transfer from completed tasks to future work.

Activation Keywords

  • ICE review
  • ICE 回顾
  • 任务回顾
  • task review
  • 知识巩固
  • consolidate knowledge
  • investigate consolidate exploit
  • 跨任务学习

Recommended Model

  • sonnet4.5 (Recommended for complex reasoning and analysis)

Tools Used

  • read: Read task history, memory files, and knowledge base
  • write: Update memory files and knowledge base
  • edit: Modify existing knowledge entries
  • memory_search: Search for related knowledge patterns
  • memory_get: Retrieve specific memory snippets

Usage Patterns

Post-Task Review

任务完成了,进行 ICE 回顾

Weekly Review

本周任务 ICE 总结

Learning Consolidation

将这段经验用 ICE 方法巩固

Instructions for Agents

Overview

The ICE strategy consists of three phases that form a continuous learning loop:

┌─────────────────────────────────────────────────────────┐
│                    ICE Review Cycle                      │
├─────────────────────────────────────────────────────────┤
│                                                          │
│   ┌─────────────┐   ┌─────────────┐   ┌─────────────┐   │
│   │ Investigate │ → │ Consolidate │ → │   Exploit   │   │
│   │   调查      │   │    巩固     │   │    利用     │   │
│   └─────────────┘   └─────────────┘   └─────────────┘   │
│         ↑                                     │          │
│         └─────────────────────────────────────┘          │
│                                                          │
└─────────────────────────────────────────────────────────┘

Phase 1: Investigate (调查)

Purpose: Systematically review the task execution process

Steps:

  1. Task Context Review

    • What was the original goal?
    • What constraints existed?
    • What resources were available?
  2. Decision Analysis

    • What key decisions were made?
    • Why were these decisions chosen?
    • Were there alternative approaches?
  3. Problem Identification

    • What obstacles were encountered?
    • How were they resolved?
    • What could have been done better?

Output Template:

## Investigate

### Task Overview
- Goal: [Original objective]
- Context: [Constraints and resources]
- Duration: [Time taken]

### Key Decisions
1. [Decision] → [Reasoning] → [Outcome]
2. [Decision] → [Reasoning] → [Outcome]

### Problems Encountered
| Problem | Solution | Effectiveness |
|---------|----------|---------------|
| [issue] | [solution] | [1-5 rating] |

Phase 2: Consolidate (巩固)

Purpose: Extract reusable knowledge and patterns

Steps:

  1. Pattern Recognition

    • Identify what worked well
    • Identify what didn't work
    • Find generalizable patterns
  2. Knowledge Abstraction

    • Convert specific solutions to general principles
    • Create reusable templates or checklists
    • Document best practices
  3. Memory Integration

    • Update MEMORY.md with key learnings
    • Create or update skill files
    • Cross-reference related knowledge

Output Template:

## Consolidate

### What Worked Well
- [Success 1]: [Why it worked]
- [Success 2]: [Why it worked]

### What Didn't Work
- [Failure 1]: [Root cause]
- [Failure 2]: [Root cause]

### Extracted Patterns
1. **[Pattern Name]**
   - When to use: [Conditions]
   - How to apply: [Steps]
   - Expected outcome: [Result]

### Knowledge Updates
- Updated: [MEMORY.md section]
- Created: [New skill/pattern]
- Cross-referenced: [Related entries]

Phase 3: Exploit (利用)

Purpose: Apply consolidated knowledge to future tasks

Steps:

  1. Identify Applicability

    • Where can this knowledge be applied?
    • What future tasks would benefit?
    • Who else might need this?
  2. Create Actionable Assets

    • Update checklists
    • Create templates
    • Configure automation
  3. Schedule Follow-up

    • Set reminders for application
    • Track effectiveness
    • Plan for iteration

Output Template:

## Exploit

### Applicable Scenarios
- [Scenario 1]: [How to apply]
- [Scenario 2]: [How to apply]

### Created Assets
- [ ] Checklist: [Name and location]
- [ ] Template: [Name and location]
- [ ] Automation: [Cron job or script]

### Follow-up Actions
- [ ] Apply in: [Next task type]
- [ ] Review effectiveness: [Date]
- [ ] Iterate: [Improvement focus]

Context Files

~/.openclaw/workspace/MEMORY.md

Long-term memory file where consolidated knowledge should be stored.

~/.openclaw/workspace/knowledge/skills/learned-skills.md

Skills learned from task experiences.

~/.openclaw/workspace/memory/YYYY-MM-DD.md

Daily notes for recording ICE reviews.

Error Handling

Incomplete Information

If task details are missing:
  1. Check session history for context
  2. Use memory_search to find related information
  3. Ask user for clarification if critical
  4. Note gaps in the review for future improvement

Conflicting Patterns

If new knowledge conflicts with existing:
  1. Document both patterns
  2. Identify conditions where each applies
  3. Create context-aware selection logic
  4. Update MEMORY.md with nuance

Examples

Example 1: Post-Task ICE Review

User: "任务完成了,进行 ICE 回顾"

Agent Process:
1. Reads recent session history
2. Identifies completed task context
3. Executes ICE phases

## Investigate

### Task Overview
- Goal: Create 3 new skills from paper learning
- Context: Skills needed to be in specific directory, with testing
- Duration: 30 minutes

### Key Decisions
1. Created skill directories first → Organized structure → Clean setup
2. Used template from existing skill → Consistent format → Fast creation

### Problems Encountered
| Problem | Solution | Effectiveness |
|---------|----------|---------------|
| Unknown directory | Asked user for correct path | 5/5 |

## Consolidate

### What Worked Well
- Checking existing skills first: Avoided duplication
- Using templates: Consistent quality

### Extracted Patterns
1. **Skill Creation Workflow**
   - When to use: Creating new skills
   - How to apply: Check existing → Ask path → Use template → Test
   - Expected outcome: Consistent, tested skill

### Knowledge Updates
- Updated: knowledge/skills/learned-skills.md
- Created: 3 new SKILL.md files

## Exploit

### Applicable Scenarios
- Future skill creation: Use same workflow
- Documentation tasks: Apply template approach

### Created Assets
- [x] Template: ICE review template in SKILL.md
- [ ] Automation: None needed

Agent: ICE 回顾完成。已更新知识库,提炼了技能创建工作流模式。

Best Practices

  1. Timely Review: Conduct ICE review within 1 hour of task completion
  2. Honest Assessment: Acknowledge failures and successes equally
  3. Specific Patterns: Extract concrete, actionable patterns
  4. Cross-Reference: Link to related knowledge in MEMORY.md
  5. Iterate: Revisit and refine patterns over time

Limitations

  • Requires clear task completion state
  • Effectiveness depends on quality of session logs
  • May miss context if task spans multiple sessions
  • Requires conscious effort to apply learnings

Resources

Related Skills

  • memory-retrieval: Two-stage memory retrieval for finding related knowledge
  • self-challenge: Self-challenge mechanism for testing skills
  • skill-extractor: Extract skills from conversations

Notes

  • ICE review is most effective when done consistently
  • The three phases form a continuous improvement cycle
  • Consider scheduling weekly ICE review sessions
  • Track the ROI of extracted patterns over time