#Prompt Templates, Variables & Chaining¶

Moving from one-off prompts to maintainable, programmatic prompt systems.

Templates With Variables¶

Never concatenate strings ad hoc. Use named placeholders:

text

8 lines

1SYSTEM: You are a {role}. Always answer in {language}.
2
3USER:
4Summarise the document for a {audience} in {max_words} words.
5
6<document>
7{document}
8</document>

Benefits: versionable, testable, reviewable in code review, and safe (you control where untrusted input is injected — never inside the instruction region).

Prompt Chaining¶

Split a complex job into a pipeline where each step's output feeds the next. More reliable than one mega-prompt (Module 2: decomposition).

Raw transcript
   │  Prompt A: extract action items (JSON)
   ▼
Action items JSON
   │  Prompt B: assign owner + due date per item
   ▼
Enriched items
   │  Prompt C: draft a polite follow-up email
   ▼
Final email

Each prompt is single-purpose, independently testable, and can use the right model size for its difficulty (cheap model for extraction, stronger for drafting — cost optimisation, Module 5).

Routing (a Special Chain)¶

A cheap classifier prompt routes the request to a specialised prompt:

text

3 lines

1Classify the request into exactly one of:
2[billing, technical, sales, other]
3Respond with one word only.

→ then dispatch to a handler prompt tuned for that category.

Versioning & Testing Prompts¶

Treat prompts like code:

Store them in source control (a prompts/ module), not scattered in strings
Give each a version id; log which version produced which output
Maintain a regression test set of input → expected-property pairs and run it whenever a prompt changes (Module 5: evaluation)

Pitfalls¶

Pitfall	Fix
Untrusted input inside the instruction block	Inject only inside fenced data region
One giant 2-page prompt doing 5 jobs	Chain into single-purpose steps
Hard-coded prompts copy-pasted everywhere	Centralised, versioned templates
No logging of prompt+version+output	Log them — you can't debug what you can't see

Takeaway: Production prompting is software engineering. Templates, chaining, versioning, and tests are not optional.

Templates With Variables¶

Never concatenate strings ad hoc. Use named placeholders:

text

8 lines

1SYSTEM: You are a {role}. Always answer in {language}.
2
3USER:
4Summarise the document for a {audience} in {max_words} words.
5
6<document>
7{document}
8</document>

Benefits: versionable, testable, reviewable in code review, and safe (you control where untrusted input is injected — never inside the instruction region).

Prompt Chaining¶

Split a complex job into a pipeline where each step's output feeds the next. More reliable than one mega-prompt (Module 2: decomposition).

Raw transcript │ Prompt A: extract action items (JSON) ▼ Action items JSON │ Prompt B: assign owner + due date per item ▼ Enriched items │ Prompt C: draft a polite follow-up email ▼ Final email

Each prompt is single-purpose, independently testable, and can use the right model size for its difficulty (cheap model for extraction, stronger for drafting — cost optimisation, Module 5).

Pitfalls¶

Pitfall	Fix
Untrusted input inside the instruction block	Inject only inside fenced data region
One giant 2-page prompt doing 5 jobs	Chain into single-purpose steps
Hard-coded prompts copy-pasted everywhere	Centralised, versioned templates
No logging of prompt+version+output	Log them — you can't debug what you can't see

Takeaway: Production prompting is software engineering. Templates, chaining, versioning, and tests are not optional.