Installable skillMIT

Search Fallback Workflow

Graceful degradation workflow for continuing tasks when web search tools fail, using internal knowledge with verification caveats

BY xyliu-csUPDATED Oct 27, 2026
INTENDED FOR✳Claude Code>_Codex◈OpenClaw↗Cursor◎ChatGPT✦GeminiInstall in a listed agent.

instructions.md

---
name: search-fallback-workflow
description: Graceful degradation workflow for continuing tasks when web search tools fail, using internal knowledge with verification caveats
---

# Search Fallback Workflow

## When to Apply

Use this workflow when:
- `search_web` tool fails repeatedly (typically 2+ consecutive failures)
- External information is needed but unavailable due to tool limitations
- Task completion is still possible with reasonable confidence using internal knowledge
- The output can include appropriate caveats about verification needs

## Core Pattern

1. **Detect Search Failure**
   - Monitor `search_web` responses for errors, empty results, or timeouts
   - After 2 consecutive failures, assess whether the task can proceed without fresh data

2. **Assess Feasibility**
   - Determine if internal knowledge is sufficient for the task at hand
   - Identify which sections/claims require external verification
   - Evaluate risk level: Can the task proceed safely with caveats?

3. **Proceed with Internal Knowledge**
   - Continue the task using well-established facts, regulations, or principles
   - Prioritize accuracy over completeness when uncertain
   - Use conservative language for claims that cannot be verified

4. **Flag for Verification**
   - Clearly mark any claims that should be verified:
     ```markdown
     [VERIFICATION NEEDED: Search tool unavailable during generation]
     ```
   - Include a verification checklist in the execution summary
   - Distinguish between high-confidence and low-confidence assertions

5. **Document Limitations**
   - Include an explicit note in the output about search unavailability
   - List specific items requiring follow-up verification
   - Provide recommendations for manual verification steps

## Implementation Template

```markdown
## Execution Notes

**Search Tool Status**: Unavailable during generation

**Confidence Levels**:
- ✓ High confidence: Established regulations, well-documented facts
- ⚠ Medium confidence: Industry standards, commonly accepted practices  
- ✗ Requires verification: Time-sensitive data, recent changes, specific citations

**Verification Checklist**:
- [ ] Verify citation: [specific reference]
- [ ] Confirm current status: [specific item]
- [ ] Review recent updates: [specific topic]
```

## Code Pattern for Agents

```python
# Pseudocode for search fallback detection
search_attempts = 0
max_search_attempts = 2
search_failed = False

for query in required_searches:
    result = search_web(query)
    if result.success:
        process_result(result)
    else:
        search_attempts += 1
        if search_attempts >= max_search_attempts:
            search_failed = True
            log_warning("Search tool unavailable, proceeding with fallback")
            break

if search_failed:
    # Apply graceful degradation
    content = generate_with_internal_knowledge()
    add_verification_flags(content)
    document_limitations_in_summary()
```

## Example Output Structure

When using this fallback, structure documents with clear sections:

```markdown
# Document Title

## Disclaimer
This document was generated with limited access to external verification tools.
Items marked with [VERIFICATION NEEDED] should be confirmed before final use.

## Content Sections
[Standard content with internal knowledge]

## Items Requiring Verification
- Topic A: Current regulatory status needs confirmation
- Topic B: Recent policy changes should be checked
- Topic C: Specific citations require source validation

## Recommendations
1. Cross-reference with official sources
2. Verify time-sensitive information
3. Consult subject matter experts for critical decisions
```

## Best Practices

1. **Don't fabricate citations** - Better to flag as unverified than invent sources
2. **Use conservative language** - "Generally," "Typically," "Commonly" instead of absolutes
3. **Prioritize safety** - If misinformation could cause harm, recommend manual verification
4. **Track what's missing** - Maintain a clear list of items needing follow-up
5. **Be transparent** - Clearly communicate limitations to end users

## When NOT to Use This Pattern

- Tasks requiring current/real-time data (stock prices, news, weather)
- Legal or medical advice requiring authoritative sources
- Compliance documentation where citations are mandatory
- Situations where incorrect information could cause significant harm

## Related Patterns

- `citation-management`: Handling references and sources in documents
- `error-recovery-workflow`: General patterns for tool failure recovery
- `confidence-annotation`: Marking certainty levels in generated content
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.