flpbalada Prompt Chaining
Break work into fixed ordered steps. Validate each handoff.
Goal
Reduce error by making each prompt do one job. Keep handoffs explicit.
Rules
- Use only for fixed sequences.
- Define final output first.
- Give each step one responsibility.
- Use strict handoff schema.
- Validate before next step.
- Stop or retry on bad output.
- Log cost, latency, input, and output when useful.
Flow
- Define final output.
- Split work into ordered steps.
- Write job and schema for each step.
- Run step 1.
- Validate output.
- Continue or retry.
- Final step formats result.
Good Step Boundaries
- Extract facts.
- Classify input.
- Summarize evidence.
- Draft answer.
- Check answer against facts.
- Format for target system.
Avoid
- Too many steps.
- Vague handoff format.
- Later step redoing earlier work.
- No validation.
- Copying full context every step.
- Using chaining when one prompt is enough.
Output
## Prompt Chain
Final output: [target]
Steps:
1. [job], input [schema], output [schema], validation [check]
2. [job], input [schema], output [schema], validation [check]
Failure rule:
- [retry / stop / escalate]
Decision Rule
Use prompt chaining when stepwise validation lowers errors more than handoff cost. If steps are independent, use parallelization.