Prompts and instructions

The prompt is the full set of instructions and information a model receives to answer a question. EDSL assembles it from question text, answer-format instructions, scenario values, agent instructions, and any prior answers supplied through piping or memory. Reading it helps you see the task from the model’s perspective—for example, whether a document passage or an earlier answer is actually present.

Inspect before execution

This example builds a classification job and renders its prompts without generating answers. Prompt cost information may depend on available model-pricing metadata.

from edsl import Agent, Model, QuestionMultipleChoice, Scenario

question = QuestionMultipleChoice(
    question_name="category",
    question_text="Classify this request: {{ scenario.request }}",
    question_options=["Billing", "Access", "Other"],
    answering_instructions="Return one of the listed options. Use Other when neither applies.",
)
agent = Agent(instruction="Apply the coding scheme literally; do not infer missing facts.")
scenario = Scenario({"record_id": "r1", "request": "Please reset my password."})
job = question.by(agent).by(scenario).by(Model())
# Rendering prompts does not generate model responses.
prompts = job.prompts()

prompts.select("user_prompt", "system_prompt")

The table contains user_prompt and system_prompt: the question-specific material and the surrounding instructions EDSL prepares for the model. Read them together with the row’s question and input identifiers. For a classification task, you should see the source text and allowed categories; reference labels kept only for later evaluation should not appear.

Change instructions deliberately

Use Agent(instruction=...) to change respondent instructions. Use a question’s answering_instructions for answer requirements, and scenario values for varying inputs. Preserve the required response format when customizing instructions. Piped answers and memory depend on earlier responses. A preview cannot show answers that have not yet been generated. Use the recorded prompts after execution to see the actual earlier answers supplied to a follow-up question. Providers may format these instructions differently, so compare prompts with the selected model in mind.

Inspect after execution

After running the job above with configured model access:

results = job.run()
results.select("prompt.*", "answer.*")

This executes model calls and may incur costs. Keep the prompts and model settings with your exported answers so that a comparison can be audited. See Prompt sensitivity for comparing alternative wording.

Common methods

Call	Arguments and result
`job.prompts(iterations=1)`	Return a Dataset of prompts for the planned interviews.
`job.show_prompts(all=False)`	Print the user and system prompts; use `all=True` to include every column, including estimated cost.
`results.select("prompt.", "answer.")`	Inspect the prompts recorded with completed responses.
`Prompt("Review {{ message }}")`	Create a reusable text template.
`prompt.template_variables()`	List the variable names referenced by the template.
`prompt.render({"message": "Export your report"})`	Return a Prompt with values substituted from the dictionary.

To try a template locally:

from edsl import Prompt

prompt = Prompt("Review this message: {{ message }}")
rendered = prompt.render({"message": "Export your report"})
rendered.text
# In a script: print(rendered.text)

Rendering this template makes no model call. Use job.prompts() to inspect the full instructions that EDSL assembles for a study. Prompt.from_txt("instructions.txt") loads prompt text from a local file. Prompt.from_template("instructions.jinja2", path_to_folder="templates", **values) loads and renders a template. Before rendering, prompt.undefined_template_variables(values) identifies missing variables and prompt.unused_traits(values) identifies supplied keys that the template does not use.

Introduction

Setup and tutorials

EDSL objects

How-to guides

Research examples

Human research

Execution and costs

Saving and sharing

Research and teaching

Expected Parrot platform

Contributing

Inspect before execution

Change instructions deliberately

Inspect after execution

Common methods

​Inspect before execution

​Change instructions deliberately

​Inspect after execution

​Common methods

Inspect before execution

Change instructions deliberately

Inspect after execution

Common methods