Prompt Chaining

Breaks a monolithic legal task into a Directed Acyclic Graph (DAG) of atomic sub-tasks. Instead of one massive prompt, the system executes a predefined sequence—e.g., (1) extract clauses, (2) classify governing law, (3) check for conflicts, (4) generate summary.

• Enforces logical order: Prevents the model from synthesizing before extracting.
• Reduces cognitive load: Each prompt focuses on a single, well-defined legal operation.
• Enables human-in-the-loop: A human reviewer can validate the output of Step 2 before the chain proceeds to Step 3.

The output of one link in the chain is programmatically injected into the prompt template of the next. This creates a stateful reasoning thread without relying on the model's internal memory.

• Explicit data passing: A contract_type variable extracted in Step 1 is passed as a parameter to the summarization prompt in Step 4.
• Prevents context collapse: Avoids the model forgetting early instructions when processing long documents.
• Supports structured formats: Intermediate outputs are often forced into JSON, ensuring the next step receives clean, machine-readable input.

Prompt chains can include programmatic gateways that route the workflow based on the content of an intermediate output. The chain is not strictly linear; it adapts to the data.

• Example: If Step 2 classifies a contract as governed by California law, the chain routes to a CA-specific analysis prompt. If Delaware law, it routes to a different branch.
• Implements legal logic trees: Encodes expert decision-making rules directly into the application's control flow.
• Reduces token waste: Only executes the relevant analytical path, avoiding unnecessary computation.

Because each step is an isolated call, prompt chaining provides a complete, step-by-step audit trail. If the final output is wrong, engineers can pinpoint the exact link where the error was introduced.

• Per-step logging: Every input, output, and latency metric for each link is recorded.
• Facilitates A/B testing: A single link in the chain can be swapped with an alternative prompt or model to measure performance impact.
• Critical for legal compliance: Demonstrates a deterministic, repeatable process to regulators, unlike a black-box single-call approach.

Prompt chaining allows for the precise insertion of external tool calls at the exact moment they are needed. The model doesn't decide when to call a tool; the chain architecture does.

• Retrieval-Augmented Generation (RAG): A chain can be designed to (1) generate a search query, (2) execute a vector database lookup, and (3) inject the retrieved case law into the final analysis prompt.
• Citation verification: A dedicated verification link can take a generated citation string and programmatically check it against a ground-truth database before the final output is shown.
• Deterministic security: Reduces the risk of prompt injection by isolating tool calls to specific, controlled stages.

A robust prompt chain implements programmatic error handling between steps. If an intermediate output fails validation, the chain can retry, fall back, or halt gracefully.

• Output validation: A step that requires a JSON output can trigger a retry if the model's response fails to parse.
• Self-Refine loops: A chain can be designed as a loop: (1) draft clause, (2) critique clause against a policy, (3) if critique fails, return to step 1.
• Graceful degradation: If a non-critical link fails, the chain can skip it and log the error, rather than crashing the entire workflow.

A foundational technique that instructs a model to generate intermediate reasoning steps before a final answer. Unlike chaining, which uses multiple API calls, CoT works within a single prompt by appending 'Let's think step by step.' For legal tasks, this improves performance on multi-hop reasoning—such as applying a statute to a novel fact pattern—by making the inferential path explicit and auditable.

A self-correcting prompting strategy designed to reduce legal hallucination. After generating an initial response, the model independently drafts a series of fact-checking questions and answers them against source material. For prompt chaining, a CoV step is often the final link, validating that every generated case citation and statutory reference is factually grounded before the output reaches the user.

An iterative strategy where the model generates a draft, critiques its own output against a rubric, and produces a revision. In legal prompt chains, a Self-Refine loop can be inserted after a summarization step to ensure the condensed text preserves deontic modalities (obligations, permissions) and does not alter the legal meaning. This transforms a linear chain into a quality-controlled feedback loop.

The mechanism by which a model emits machine-readable JSON or triggers an API. Prompt chains rely on structured output at each step to ensure the next link receives a predictable data schema. For example, a contract extraction step must output a JSON object with clause_type and text keys so a downstream classification step can parse it deterministically without fragile regex parsing.