Workflow patterns

The patternsCopy Anchor Link

Prompt chaining: break a task into steps where each output feeds the next, with checks between. Routing: classify the input and send it to a specialized prompt or model. Parallelization: run independent subtasks at once, or run the same task several times and vote. Orchestrator-workers: a lead model splits an unpredictable task and delegates pieces to workers. Evaluator-optimizer: one model drafts, another critiques, and the loop repeats until the criteria are met.

ChoosingCopy Anchor Link

Chain when the task splits cleanly. Route when inputs differ in kind. Parallelize for speed or confidence. Use orchestrator-workers when subtasks are not known beforehand. Use evaluator-optimizer when you have clear criteria and revision helps. Each pattern adds calls, so add one only when it improves a measured result.

A worked example: prompt chainingCopy Anchor Link

Task: write a product page. Step 1: extract key facts from the spec as bullets. Step 2: draft copy from the bullets. Step 3: check the copy against the bullets and list any claim not supported. Step 4: fix those. Each step is small and checkable, so errors are caught early.

A worked example: routingCopy Anchor Link

An incoming message is classified as billing, technical or sales. Each category goes to a prompt (or model) tuned for it. A cheap model does the classification, a stronger one handles the technical cases.

Parallelization with a voteCopy Anchor Link

Ask the model three times to label a message. The labels come back bug, bug, billing. The majority is bug. That costs three calls instead of one.

Invented math: if one call is right 80 percent of the time and the three runs make independent mistakes, the majority is right when at least two are right. That is 3 × 0.8 × 0.8 × 0.2 + 0.8³ = 0.384 + 0.512 = 0.896, about 90 percent. The catch is the word independent. If the model misreads a kind of message the same way every time, three votes give the same wrong answer.

How evaluator-optimizer stopsCopy Anchor Link

The loop needs an exit. A common version is: draft, then a critic checks the draft against written criteria, then the writer revises. It stops when every criterion passes or after three rounds, whichever comes first.

Two failures to watch for. A critic with vague criteria always finds something to improve, so the loop runs to the cap. Or the revisions flip back and forth between two versions. Confirm either one by saving every round and comparing round 2 with round 4. If they are near identical, cut the loop and fix the criteria.

When one prompt beats all of themCopy Anchor Link

Before adding a pattern, run the single plain call on your test set. If it gets 90 of 100 right, a three-step chain has to beat 90 to justify three times the calls. Often the better fix is a clearer prompt, not more structure.

Common mistakeCopy Anchor Link

Adding patterns without measuring. Each extra call adds cost, time and failure points.