Society · Protocols
PromptChain
Sequential LLM calls with validation gates
Definition
Mechanism
Decomposes a Task into a fixed Sequence of LLM calls. Each step's output is validated against an AcceptSpec before it is passed to the next, and a failed output is never forwarded. After a failed validation, the chain invokes ReAttempt on that step up to `max_retries_per_step`; when that budget is exhausted, it halts without executing the next step. A value of zero halts on the first failed validation. It orchestrates ToolInvoke calls and applies InputGuard validation at each boundary.
Risks
Failure modes
Undetected error propagation
a flawed intermediate output passes validation and corrupts the next step.
Validation brittleness
an overly narrow AcceptSpec rejects a valid intermediate output.
Relationships
Builds on
Accepts
Infrastructure
The atomic unit of intent
Composes with
Infrastructure
Input validation and sanitization
Infrastructure
Substrate-level re-attempt: same call, after a delay
Infrastructure
Execute external tool and observe result
References
Infrastructure
Non-compensatory acceptance contract; consumer Verify yields Solution or FrameError
Infrastructure
Sequential data container (linked list)
Infrastructure
Sequential ordering
Used by
References
Mind
Building complex behavior from simple primitives
Mind
Use prompts to generate or refine other prompts
Reference
Parameters
max_retries_per_stepInteger, [0, 3]- Re-attempts allowed after a failed step; 0 halts on the first failure