S1
The Engineering Scope Ladder
Learn the five scopes of LLM engineering and why each new stage exists.
- 01Name the five engineering stages from prompt to graph
- 02Explain what control each stage adds
- 03Choose the simplest stage that can solve a task
One model, five engineering scopes
orientation- Prompt engineering shapes one request.
- Conversation engineering manages turns and context.
- Response engineering defines what a valid answer looks like.
- Loop engineering repeats model and tool work with budgets and checks.
- Graph engineering makes branches, joins, cycles, state, and human gates explicit.
Prompt, Conversation, Response, Loop, and Graph engineering appear in order. Each stage answers a larger control question and keeps the earlier stages inside it.
Text description
Prompt, Conversation, Response, Loop, and Graph engineering appear in order. Each stage answers a larger control question and keeps the earlier stages inside it.
A worked example: extract one invoice
Suppose you receive one invoice and need the vendor, total, currency, and due date.
- Start with prompt engineering: write a clear extraction instruction and include one example.
- If a user corrects the invoice over several messages, add conversation engineering.
- If code must consume the result, add response engineering with a schema.
- If missing fields require lookup tools and retries, add loop engineering.
- If several documents can be checked in parallel and require an approval step, add graph engineering.
The task did not become “more agentic.” Its control requirements expanded.
Common MisconceptionGraph engineering replaces prompt engineering.
CorrectionEvery graph node still needs good prompts and response contracts. Larger stages contain the smaller ones.
Common MisconceptionMore agents means more capability.
CorrectionExtra agents help only when specialization or parallel work outweighs coordination cost.
Deep dive: Workflow vs. agent
A workflow follows control flow defined mostly by code: run A, then B, then choose C or D. An agent gives the model more authority to choose actions or the next step.
Both can use graphs. A graph is a way to make control flow explicit; it does not automatically mean “multi-agent.”
References & deeper reading (2)
- Building Effective AgentsAnthropic Engineering
- Agents GuideOpenAI Developers
Retrieval Practice
Check one idea at a time
Put the engineering stages in order from smallest to largest control scope.
Put these in the correct order.