research-reflection

Files (1)

• SKILL.md

SKILL.md (first lines)

# Research Reflection (Reflexion Pattern)

Self-evaluate completed research tasks, store lessons learned, and retrieve relevant past experience to improve future research quality. Based on the Reflexion framework (verbal reinforcement learning without weight updates).

## When to Use

- After completing any substantial research task (literature review, data analysis, systematic review)
- Before starting a new task in a previously explored domain
- When a research approach failed and you need to diagnose why
- "What did I learn from the last CRISPR review?"
- "Before starting, check if we've done similar work before"

## When NOT to Use

- During active research execution (focus on the task)
- For simple factual lookups
- For non-research tasks

## Reflexion Cycle

The cycle runs after each substantial research task:

```
Execute Task → Evaluate Output → Generate Reflection → Store in Memory → Retrieve Next Time
```

### Phase 1: Self-Evaluation

After completing a research task, evaluate against these dimensions:

```markdown
## Post-Task Evaluation

### Completeness (1-5)
- Did I search enough databases? Which ones did I miss?
- Did I follow all 6 Research Depth Phases from SCIENCE.md?
- Are there obvious gaps in my coverage?

### Accuracy (1-5)
- Did every citation come from a tool result?
- Did I cross-verify claims against primary databases?
- Were statistical results correctly reported?

### Efficiency (1-5)
- Did I use the right tools for each subtask?
- Were there redundant searches I could have avoided?
- Did I hit rate limits that slowed me down?

### Depth (1-5)
- Did I go beyond surface-level results?
- Did I trace citation chains?
- Did I identify contradictions and open questions?

### Actionability (1-5)
- Is the output directly useful to the user?
- Are next steps clearly identified?
- Is the methodology reproducible?
```

### Phase 2: Generate Reflection

Produce a structured reflection:

```markdown
## Reflection: [Task Title]
Date: [ISO date]
Domain: [e.g., molecular biology, econometrics]
Task type: [literature review / data analysis / systematic review / ...]

### What went well
- [Specific successful strategies]

### What went poorly
- [Specific failures or inefficiencies]

### Key lessons
- [Actionable insights for future tasks]

### Tools that worked best
- [Tool name]: [Why it was effective for this task]

### Tools that underperformed
- [Tool name]: [Why, and what to use instead]

### Domain-specific insights
- [API endpoints that were useful]
- [Search strategies that worked]
- [Data sources that were reliable/unreliable]

### Score: [total/25]
```

### Phase 3: Store Reflection

Store reflections using the science-memory extension:

```bash
# Store reflection with domain tags
curl -X POST http://localhost:18789/memory/store \
  -H "Content-Type: application/json" \
  -d '{
    "key": "reflection-[task-id]",
    "content": "[reflection markdown]",
    "tags": ["reflection", "domain-tag", "task-type"],
    "metadata": {"score": 18, "date": "2026-03-11"}
  }'
```

### Phase 4: Retrieve Before New Task

Before starting a new research task, retrieve relevant past reflections:

```bash
# Search for reflections in the same domain
curl http://localhost:18789/memory/search?q=domain-tag+task-type&tags=reflection
```

Use retrieved reflections to:
1. Avoid repeating past mistakes

From beita6969/ScienceClaw, MIT.

More from beita6969/ScienceClaw