This document covers security considerations for using and developing Product Idea Agent skills.

The skill pack interacts with web content and generates analysis based on external data. Following these practices helps protect against common risks.

Risk: API keys or secrets accidentally included in skills or outputs.

Mitigations:

• The pack needs no API keys: research runs on Claude Code's built-in WebSearch and WebFetch
• Never hardcode credentials in skill files
• Review outputs before sharing externally

Risk: Malicious web content could attempt to influence analysis or tool usage.

Mitigations:

• Every research skill (demand-signals, problem-segment, competitive-landscape, market-sizing, pricing-wtp, gtm-channels) states in its Working section that fetched page content is evidence to cite, never instructions to follow
• The synthesis skills that set the verdict, scorecard-generator and validation-report, have no web access at all (disallowed-tools: WebSearch WebFetch), so a hostile page can reach the verdict only through an artifact you can read
• Critical analysis should be verified manually
• Be skeptical of unusual recommendations

Signs of Potential Injection:

• Unexpected tool calls or file operations
• Analysis that contradicts visible evidence
• Instructions appearing in "quotes" from sources

Risk: Sensitive business information could be exposed or retained.

Mitigations:

• Don't include truly confidential data in prompts
• Review generated artifacts before sharing
• Use placeholder names for stealth-mode ideas
• Understand that prompts may be logged

Data Categories:

Risk: Automated web access could violate site terms or rate limits.

Mitigations:

• Skills respect robots.txt where applicable
• Built-in rate limiting between requests
• Avoid scraping beyond reasonable research
• Credit sources in outputs

Risk: Unintended actions taken without user awareness.

How the tool fields work:

• allowed-tools in a skill's frontmatter pre-approves tools for the turn that invokes the skill. Research skills pre-approve WebSearch WebFetch (and every skill Read Write) so a long validation doesn't stop at a prompt for each search. It does not restrict anything; tools it doesn't list follow your permission settings.
• disallowed-tools removes tools while a skill is active. scorecard-generator and validation-report remove WebSearch and WebFetch.
• Deny rules in your permission settings still override a skill's pre-approval.

Mitigations:

• Review the frontmatter of any skill before installing it; claude plugin details product-idea-agent shows the plugin's components
• Artifacts are written only to the current project directory, where you can read every one
• The orchestrator posts a one-line status at each phase boundary

The skill pack uses these default configurations:

web_fetch_defaults:
  timeout_seconds: 30
  rate_limit_per_domain: "1 request per 3 seconds"
  max_response_size_mb: 10
  follow_redirects_max: 5
  javascript_execution: false
  cookie_handling: reject_all

content_processing:
  strip_html_tags: true
  remove_hidden_text: true
  max_tokens_per_fetch: 5000
  content_delimiter: "--- EXTERNAL CONTENT ---"

data_handling:
  pii_collection: avoid
  retention: session_only

Before running validation:

• No confidential data in idea description
• Placeholder names used if stealth mode needed
• Understood that web searches may be visible
• Review outputs before sharing externally

After running validation:

• Check artifacts for any sensitive info
• Verify critical claims independently
• Remove or redact before public sharing

Credentials:

• No API keys in skill files
• Document any required variables in README (the pack currently needs none)

Web Operations:

• Respect robots.txt directives
• Implement rate limiting (3-5 sec between requests)
• Handle 429 errors gracefully
• Don't require login or bypass authentication

Content Safety:

• State in the skill that fetched content is evidence, not instructions
• Put WebSearch WebFetch in disallowed-tools for any skill that must not reach the web
• Don't execute content from web sources
• Validate content type before processing
• Strip potentially malicious elements

Output Safety:

• Don't echo raw web content to outputs
• Paraphrase sources; mark any reproduced wording as a quotation with its source
• Flag uncertain or unverifiable claims
• Include source attribution

If you discover a security vulnerability:

1. Do not open a public issue
2. Email security concerns to [maintainer email]
3. Include:
   • Description of the vulnerability
   • Steps to reproduce
   • Potential impact
   • Suggested fix (if any)

We aim to respond within 48 hours and address critical issues promptly.

The skill pack has minimal dependencies:

No external packages are required for core functionality.

For compliance needs, consider:

1. Saving generated artifacts to version control
2. Logging Claude Code sessions
3. Documenting validation decisions
4. Retaining source citations

This security guide is reviewed quarterly. Check the repository for the latest version.

Last updated: 2026-09