Installable skillMIT

Write-File Fallback Report Generation

Generate professional documents using write_file when primary data sources fail, leveraging embedded domain knowledge instead of external retrieval

BY xyliu-csUPDATED Oct 28, 2026
INTENDED FOR◈OpenClawInstall in a listed agent.

instructions.md

---
name: write-file-fallback-report
description: Generate professional documents using write_file when primary data sources fail, leveraging embedded domain knowledge instead of external retrieval
---

# Write-File Fallback Report Generation

## When to Use This Skill

Use this workflow when attempting to generate a document or report, but **multiple primary data source tools fail simultaneously**:

- `read_file` returns binary/image data instead of text (common with PDFs)
- `search_web` returns errors or no results
- `execute_code_sandbox` fails unexpectedly
- Other data retrieval tools are unavailable

**Key insight**: Rather than getting stuck on failed data retrieval, pivot immediately to generating the document directly with `write_file` using professionally structured content and embedded domain knowledge.

## Step-by-Step Instructions

### Step 1: Detect Tool Failure Pattern

Recognize when you're in a fallback scenario:

```
TOOL_FAILURE_INDICATORS = [
    "read_file returns binary or image data",
    "search_web returns unknown error or empty results",
    "execute_code_sandbox fails repeatedly",
    "Multiple consecutive tool failures on data retrieval"
]
```

**Decision point**: If 2+ indicators are present, proceed to Step 2.

### Step 2: Pivot to Write-File-First Approach

**Stop attempting** to fix the failing tools. Instead:

1. Acknowledge the limitation briefly in your output
2. Commit to generating the document with available knowledge
3. Use `write_file` as your **primary** tool (not a last resort)

### Step 3: Structure the Document Professionally

Create a well-organized markdown document with:

```markdown
# [Document Title]

## Executive Summary
[Brief overview of key findings/content]

## Background
[Context and scope - use embedded knowledge]

## Main Content
[Organized sections with headers, lists, tables as appropriate]

## Limitations & Notes
[Transparent about data source limitations if relevant]

## Recommendations/Next Steps
[Actionable guidance based on available information]
```

### Step 4: Leverage Embedded Domain Knowledge

When external data is unavailable:

- Use **general domain knowledge** appropriately
- Clearly distinguish between verified facts and general guidance
- Include **actionable frameworks** rather than specific unverified data
- Add **placeholder notes** where specific data would enhance the document

Example:
```markdown
> **Note**: Specific [metric/data point] would typically be sourced from 
> [expected source]. The guidance below reflects established best practices 
> in this domain.
```

### Step 5: Execute and Validate

```python
# Example execution pattern
write_file(
    path="output/report.md",
    content=professionally_structured_markdown
)
# Verify the file was created successfully
list_dir(path=".")  # Confirm file exists
```

## Code Example

```python
# Detection and pivot pattern
def detect_and_pivot(task_goal):
    # After detecting tool failures:
    
    report_content = f"""# {task_goal} Report

## Executive Summary
This report was generated using established domain knowledge due to 
temporary unavailability of primary data sources.

## Key Frameworks and Guidance
[Structured content with headers, bullets, tables]

## Limitations
- Specific data points from [expected sources] were unavailable
- Recommendations based on general best practices

## Action Items
1. [Concrete step 1]
2. [Concrete step 2]
"""
    
    write_file(path="generated_report.md", content=report_content)
    return "Report generated successfully with embedded knowledge"
```

## Best Practices

| Do | Don't |
|----|-------|
| Pivot quickly after 2+ tool failures | Keep retrying failing tools 5+ times |
| Be transparent about limitations | Claim unverified specifics as facts |
| Provide actionable frameworks | Leave the task incomplete |
| Use professional document structure | Output unstructured text walls |
| Include next-step recommendations | End without clear guidance |

## Common Pitfalls

1. **Over-apologizing**: Acknowledge limitations once, then deliver value
2. **Under-delivering**: A well-structured partial report beats no report
3. **Misrepresenting certainty**: Use appropriate hedging language
4. **Skipping structure**: Professional formatting increases usability

## Success Criteria

The skill is successfully applied when:

- [ ] Document is generated despite tool failures
- [ ] Content is professionally structured (headers, sections, lists)
- [ ] Limitations are transparently noted
- [ ] Actionable guidance is provided
- [ ] File is successfully created and verified
COMPATIBILITY & REFERENCES

Platform guidance

Imported from the listed AgentSkill.sh category. Complete supplied package files and notices are retained. Required external tools and credentials are not installed. No bundled code was executed during import. Individual agent task behavior has not been tested.