Skip to main content
Your Expert isn’t static. After publishing, the Darwin evolution system continuously analyzes real user conversations and automatically generates skill improvement suggestions. All you need to do is periodically review and decide whether to adopt them.
Think of it as a tireless product manager helping you iterate — it discovers improvement opportunities from real user behavior, and you make the final call.

​
How Evolution Works

The Darwin system runs automatically in the background, completely transparent to you:
  1. Data Collection: When users converse with your Expert, the system records interaction patterns
  2. Pattern Analysis: Once sufficient conversation volume is reached, the system analyzes which scenarios could be improved
  3. Suggestion Generation: Based on the analysis, specific skill improvement suggestions are generated
  4. Awaiting Review: Suggestions appear in your Studio evolution panel, waiting for your decision
The frequency of evolution suggestions depends on your Expert’s actual usage volume. Higher usage means more evolution opportunities.

​
Viewing Evolution Suggestions

1

Open Studio

Navigate to your Expert’s editing page.
2

Switch to the 'Evolution' Tab

Here you’ll find all pending evolution suggestions.
3

Review Improvement Details

Each suggestion includes: a before vs. after skill comparison, along with quality score changes.
4

Make Your Decision

For each suggestion, you can choose to “Accept” or “Ignore.”

​
Suggestion Structure

Each evolution suggestion contains the following information:

​
Dual-Ratchet Rule

The Darwin system follows a strict “dual-ratchet” rule to ensure evolution moves in the right direction:
Quality must not decrease AND efficiency must improve — this is the hard constraint on evolution suggestions.
Specifically:
  • Quality group metrics cannot decline (e.g., professional accuracy, expression clarity)
  • Efficiency group metrics must strictly improve (e.g., response conciseness, resource usage efficiency)
  • The two groups cannot offset each other — “sacrificing quality for efficiency” is not allowed
This means every suggestion you see is guaranteed not to make your Expert “dumber” or “worse.” It will only become more efficient while maintaining quality.

​
Accepting Suggestions

When you click “Accept”:
  • The improvement is immediately applied to the corresponding skill
  • Changes take effect for all subsequent conversations
  • No re-publishing or re-review is needed (skill improvements are real-time internal optimizations)
  • The operation is recorded in the evolution log
We recommend checking the evolution panel at least once a week. The more suggestions accumulate, the more efficient a batch review becomes.

​
Ignoring Suggestions

When you click “Ignore”:
  • The suggestion is marked as handled
  • No changes are made to your Expert
  • Ignored suggestions won’t reappear
If you’re unsure whether a suggestion is good, feel free to ignore it. If the direction truly has value, similar suggestions will emerge again in the future (based on new conversation data).

​
Evolution Statistics

In the evolution panel, you can also view:
  • Total Suggestions: How many suggestions the system has generated for you
  • Acceptance Rate: What percentage of suggestions you’ve adopted
  • Quality Trend: Your Expert’s quality curve over time
  • Efficiency Trend: The improvement trajectory of efficiency metrics
This data helps you intuitively feel your Expert “growing.”

​
Evolution vs. Memory

The two are independent yet complementary: memory lets the Expert understand each user’s personalized needs, while evolution continuously improves the Expert’s overall capabilities.

​
Best Practices

We recommend spending 10 minutes per week browsing evolution suggestions. Letting too many pile up not only wastes the value the system generates, but may also result in outdated suggestions mixed in.
If a suggestion significantly improves the quality score, it’s usually worth accepting. If the improvement is negligible, decide based on your judgment.
Darwin is based on real data analysis, but it doesn’t understand your “product positioning.” If a suggestion is technically sound but diverges from your intended direction, ignore it decisively.
After accepting a few suggestions, watch the usage data over the following days. Increasing conversation turns, more positive feedback, and improved retention are all positive signals.

​
Next Steps

Analytics

Measure the impact of evolution with data

Promo Notes

Actively promote your Expert