JevMade field notes / Technical guide

Try completion rules before letting them interrupt an AI assistant

limpet compares an assistant's proposed final reply with written completion rules. It recommends observing decisions before enabling interruptions, then allows only one extra attempt when a rule triggers.

Original by noplan-incAgent workflows

Listen to this guide

JevMade’s plain-English explanation

0:00 /

AI narration

Credits and license

“limpet” by noplan-inc. Read the original source.

This expanded guide is an AI-narrated adaptation of the source’s essential explanation, examples and caveats, not a word-for-word reading. The synthetic voice does not imitate the author or imply their endorsement.

License: MIT

MIT License

Copyright (c) 2026 no plan inc.

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

Our summary

An AI assistant may stop before satisfying a user's completion rule. limpet checks the proposed final reply against rules written in ordinary language. When an enabled rule triggers, it sends the assistant back once with a reminder, then allows the next stop to avoid an endless loop.

The project recommends first recording judgments without blocking anything. Tools examine earlier conversations and estimate which rules distinguish wanted from unwanted stops. They choose cutoffs while limiting mistaken interruptions, and weak rules remain visible without automatically becoming active blocks.

The reported evaluation uses one author's conversations and imperfect labels produced by a model. Even the strongest rules catch only some unwanted stops, and later evidence may reveal problems the final reply cannot show. Conversations leave the computer, and a reminder is not proof that work was verified.

Key takeaways

Observe proposed interruptions before enabling them.
Tune rules on your own examples rather than copying cutoffs.
Limit extra attempts so a mistaken judgment cannot trap the assistant.

More guides →See the experiments →