Standard AI agents are great at listing options but terrible at identifying the actual root conflict in complex engineering problems. A new developer tool called the TRIZ skill attempts to fix this by teaching agents Genrich Altshuller’s 1940s Soviet invention method, forcing them to name contradictions before proposing solutions. The result, according to a recent DEV.to post by user viiko, is a dramatic improvement in plan quality, with blind evaluations showing agents acting on skill-enhanced plans 38 out of 39 times compared to just once for the baseline.

The Soviet Secret to AI Planning

TRIZ, or Theory of Inventive Problem Solving, was developed by Soviet patent examiner Genrich Altshuller, who studied thousands of patents to identify recurring patterns in innovation. He determined that hard engineering problems are essentially contradictions: improving one parameter inevitably worsens another. The new agent skill packages this logic into a structured six-stage process, guiding the AI from problem intake through to a validated plan. By mapping specific issues to Altshuller’s 39 engineering parameters and 40 inventive principles, the agent avoids generic advice like 'add caching' and instead identifies structural fixes, such as separating functions or changing parameters dynamically.

Blind Evaluations Show Massive Gains

The developer conducted a rigorous blind evaluation across 13 different problem cases, running three iterations per case for both the standard agent and the TRIZ-enhanced version. The data was stark: while both arms often identified similar core fixes, the TRIZ skill consistently delivered actionable plans with dated targets, measurable risk signals, and clear resource costs. The baseline agent produced vague suggestions, whereas the skill-enforced agent provided specific thresholds, such as 'wrong-answer rate above 4% for a week,' making it far more useful for actual project management. The median number of user replies required also increased, indicating the skill forces more interactive clarification rather than dumping a generic list.

Key Takeaways

  • The TRIZ skill uses a 39x39 contradiction matrix to map engineering trade-offs to proven inventive principles.
  • Blind grading showed a 38/39 preference for TRIZ-enhanced plans over baseline outputs for actionability.
  • The skill enforces specific fields for 'Resources,' 'Watch for,' and 'Done when,' ensuring plans are measurable.
  • Initial failures in autonomous mode were resolved by removing word count caps and adding metric breakdowns.
  • The tool is available via a Claude Code plugin or as a downloadable zip for other agent frameworks.

The Bottom Line

This is a masterclass in prompt engineering moving beyond vibes to structured logic. If you want your agents to stop offering compromises and start solving contradictions, this skill is the missing link.