Stop sequences and delimiters

About

Stop Sequences and Delimiters are techniques used to control where the model stops generating output and how different parts of the prompt/output are clearly separated.

They work at two levels:

• Stop Sequences → Define when generation should stop

• Delimiters → Define boundaries between sections (context, instructions, output)

Core idea:

Clearly define boundaries — where input ends, sections change, and output must stop.

These techniques help prevent:

• Over-generation

• Mixing of sections

• Unstructured responses

• Leakage of unwanted text

Why Stop Sequences and Delimiters Are Critical

Without clear boundaries, models may:

• Continue generating unnecessary text

• Blend instructions with output

• Include context in the response

• Produce inconsistent endings

In engineering systems, this leads to:

• Parsing failures

• Invalid outputs (especially JSON)

• Hard-to-control responses

• Unexpected behavior

Stop sequences and delimiters ensure:

• Clean separation of sections

• Controlled output termination

• Predictable response boundaries

The Purpose of Stop Sequences and Delimiters

These techniques aim to:

1. Control output length and termination

2. Prevent unwanted text generation

3. Clearly separate context, instructions, and tasks

4. Improve parsing and reliability

5. Enforce strict response boundaries

They act as guardrails around generation.

Where They Fit in the Prompt Lifecycle

Problem Definition
      ↓
Input Design (Use delimiters to separate sections)
      ↓
Reasoning
      ↓
Knowledge Grounding
      ↓
Output Generation
      ↓
Stop Sequences Applied  ← (Terminate output cleanly)

• Delimiters → applied during input design

• Stop sequences → applied during output generation

Different Patterns

1. Section Delimiters

Used to separate parts of the prompt:

Context: <<< ...

Task: ...

Benefits:

• Prevents mixing of context and instructions

• Improves clarity

2. Instruction Delimiters

Clearly separate instructions from data:

INSTRUCTIONS:

...

DATA:

...

Helps the model distinguish intent vs input.

3. Output Delimiters

Force output within boundaries:

START_OUTPUT ... END_OUTPUT

Useful for:

• Parsing

• Preventing extra text

4. Stop Sequences (Hard Stop)

Define exact token/phrase where generation must stop.

Example:

Stop sequence: END_OUTPUT

Model stops when this appears.

5. Multi-Section Delimiter Patterns

For complex prompts:

• Context block

• Rules block

• Task block

Each clearly separated.

Improves reliability in large prompts.

Common Mistakes

1. Not Using Delimiters in Complex Prompts

Without boundaries:

• Model may confuse context with instructions

• Output may include unwanted sections

2. Weak or Ambiguous Delimiters

Using common words like:

• “Start” / “End”

can confuse the model.

Prefer unique markers:

• <<< >>>

• ---

• ##

3. Not Enforcing Output Boundaries

If you don’t define:

• Where output starts/ends

Model may:

• Add explanations

• Continue beyond expected output

4. Ignoring Stop Sequences in API Usage

Even with good prompts:

• Model may over-generate

Stop sequences at API level are critical for strict control.

5. Mixing Delimiters and Content

If delimiters appear inside actual data:

• Model may misinterpret boundaries

Choose delimiters that do not conflict with content.

Sample Prompts

Without Delimiters

Use the following data to validate the transaction.

Transaction data:
{id: 1, amount: 100}

Validate and return result.

Issues:

• Context and task not clearly separated

• Output may include extra explanation

With Delimiters and Stop Control

INSTRUCTIONS:
---
Validate the transaction and return result in JSON.
Do not include any text outside JSON.
---

DATA:
<<<
{id: 1, amount: 100}
>>>

OUTPUT FORMAT:
START_OUTPUT
{
  "isValid": boolean
}
END_OUTPUT

Benefits:

• Clear section separation

• Controlled output boundaries

• Easier parsing

• Reduced ambiguity