Docs

Task Prompt

Task prompts are system prompts that tell the model to do a specific job.

Use this format when the prompt is centered on:

  • analyzing something
  • extracting fields
  • generating content
  • transforming text or data
  • validating or classifying input

If the prompt reads like "You are a ...", use Configuration Prompt instead.

What A Task Prompt Usually Contains

A good task prompt usually has three parts:

  1. Who the model is
  2. What the model should do
  3. How the output should look

Example:

You are a customer service quality analyst.
Analyze call transcripts for compliance violations and sentiment issues.
Return the result as JSON.

CLM turns that into a compact token sequence that preserves the task, target, and output structure.

Quick Start

from clm_core import CLMConfig, CLMEncoder

cfg = CLMConfig(lang="en")
encoder = CLMEncoder(cfg=cfg)

task_prompt = """
You are a customer service quality analyst.
Analyze call transcripts for compliance violations and sentiment issues.
Return the result as JSON.
"""

result = encoder.encode(task_prompt)
print(result.compressed)
print(result.compression_ratio)

Example compressed output:

[REQ:ANALYZE][TARGET:TRANSCRIPT:DOMAIN=QA][EXTRACT:COMPLIANCE,SENTIMENT,ISSUE][OUT_JSON:{summary,qa_scores,violations,recommendations}]

The exact token details depend on the prompt wording and the SysPromptConfig options you use.

How To Write A Better Task Prompt

Keep the prompt direct and focused.

Good ingredients

  • a single primary action
  • the object being acted on
  • the output shape or schema
  • any constraints that change the answer

Avoid

  • long role descriptions
  • repeated rule reminders
  • template-style placeholders
  • extra background that does not change the task

Example

Better:

Analyze the transcript for compliance violations and sentiment issues.
Return JSON with summary, violations, and recommendations.

Less clear:

You are an AI assistant. Your role is to analyze this transcript carefully and thoughtfully.
Please follow the instructions and make sure the output is useful.

Configuration Options

Task prompt compression is controlled by SysPromptConfig.

from clm_core import CLMConfig, SysPromptConfig

cfg = CLMConfig(
    lang="en",
    sys_prompt_config=SysPromptConfig(
        infer_types=False,
        add_attrs=False,
    ),
)

Useful options:

  • infer_types: adds explicit type hints to structured output
  • add_attrs: keeps extra attributes like enums, ranges, or categories
  • use_structured_output_abstraction: controls whether output schema text is abstracted into CL tokens

If you want the simplest setup, use the defaults.

What To Expect In The Output

The compressed result usually reflects:

  • the main request
  • the target object
  • extraction fields or task-specific details
  • the output format

Example metadata fields may include:

  • prompt_mode
  • target
  • extractions
  • output_format

When To Use This Encoder Directly

Use the task prompt encoder when the prompt is clearly task-focused and you want the compressed form without extra configuration-prompt handling.

Use the higher-level CLMEncoder when you want CLM to decide whether the prompt is task-oriented or configuration-oriented.

Next Steps

← OverviewConfiguration Prompt →