We use the term personality engine to describe our Twitter bots.
The bots are built to reply based upon Twitter searches (thank you, Summize). The bots are built to URL-shorten where possible and appropriate (thank you, bit.ly and your pufferfish). The bots will be enhanced to direct message once there is a need to do so.
There are very cheap and dirty algorithms behind the bot personalities, all based upon proportional randomness. Some recurring principles are:
- Infrequent re-use of tweets, with long-term tracking history
- Functional repeats allowed but constrained by decreased proportional likelihood
- Random sentence generation, with multiple contextual grammars
- Self-exhausting ratio pools
You can tell we're not pure Statisticians. It's all a gut feel.
The most important aesthetic that the personality engines attempt to maintain is followability. Tweet timing is intended to be somewhat sparse, so that when you choose to actually follow the bot, it's a sporadic and entertaining experience vs. a bullrush of repetetive crap. It won't shock you when the bot tweets, but it should not overwhelm.
You opinions are very welcome. Please contact us if you have suggestions. There is a highly-configurable platform running behind these bots, so aesthetic changes are reasonably straightforward, and not everything has to be crafted from scratch. However, building an engine that exhibits some sort of personality is not a quick turn-around project :) and as such it requires inspiration.