AI tuning

Added · Seed

The progressive shaping of an agent’s behaviour to match the actor’s evolving needs, transforming a generic capability into a personalised collaborator.

Filters

A correction that keeps recurring is escalated into a standing rule — a trigger, the exceptions to it, and the constraints on what the agent may do in response. Composing that rule, and checking what it will do before saving it, is rule composition; the escalation from the per-case exception (“never this one again”) to the general rule is described there too.

…

Parameters

…

Voice and tone

…

References

…

Workflow

…

Consequences

  • the standing behaviour is the actor's to shape, so a correction made once holds across sessions instead of being re-issued each time
  • how far the system may act on its own is now something the actor set, which moves the responsibility for that boundary onto them
  • tuning accumulates and then goes quiet: behaviour configured months ago reads later as simply how the system is, including to the actor who configured it

Resources & references

Appears in

Delegating work

Related patterns

Serves

  • Delegation — shaping initial behaviours at delegation's set-up, and repairing misfires during exception handling

Enacts

  • Agency — the actor sets how far the system can act on its own, then lets it run

Complements

  • Localization — locale as a generation parameter
  • Generated content — settings that influence how content is generated and presented.
  • Selective memory — remembered facts and shaped behaviour are two channels of the same persistence
  • Settings — specialized settings for configuring AI behaviour
  • Workflow — a workflow builder is a tuning tool: editing the procedure is one of the ways an agent's behaviour gets progressively shaped

Related to

Preceded by

  • Agent — the agent's degree of freedom is a setting rather than a fixed property; tuning is the move that acts on it
  • Prompt — lets the actor refine how the model gets prompted

Enabled by

  • Rule composition — the tuning that survives the session is written as a rule, which is the form the correction takes