AIAgent Development
ADLC · Practice Area 08

Agent Development & Prompt Engineering

Prompt engineering is both a writing discipline and an operational discipline: define who the agent is, what it should do, the boundaries it must respect, and then manage those prompts like production code.

System prompt designAgent instructionsBehavioral constraintsPersona designPrompt versioningPrompt testing
Agent Development and Prompt Engineering title slide
Prompt craft sets behavior; Prompt-Ops keeps that behavior controlled over time.
Overview

From prompt craft to Prompt-Ops

The deck divides the practice into six levers. The first four shape the prompt itself; the last two turn prompts into versioned, tested infrastructure.

1

System Prompt Design

Set the agent's role, stable context, ground rules, capabilities, tools, and output format.

2

Agent Instructions

Tell the agent precisely what to do, how to do it, and what successful completion means.

3

Behavioral Constraints

Define the lines the agent must not cross and back prompt rules with code-level enforcement.

4

Persona Design

Shape a consistent voice and character without allowing persona to replace real constraints.

5

Prompt Versioning

Track immutable versions, diffs, promotions, audit history, and fast rollback.

6

Prompt Testing

Run each prompt version against a dataset, score outputs, compare with a baseline, then ship or iterate.

Treat prompts like code. Small wording changes can shift behavior, so review diffs, test every change, and make rollback routine.
Craft

System Prompt Design

The system prompt is the agent's charter: stable, top-level context that frames every response.

  • Define role and identity, domain context, capabilities and tools.
  • State rules and guardrails clearly, then specify output format.
  • Keep dynamic data downstream and keep the system prompt lean.
  • Use structure the model can parse, such as clear sections or XML tags where appropriate.
System Prompt Design slide showing prompt anatomy and design principles
Craft

Agent Instructions

Ambiguous instructions force the model to fill gaps with guesses. Good instructions reduce that space for interpretation.

  • Be specific and unambiguous; prefer positive framing.
  • Specify fields, length, structure, ordering, and success criteria.
  • For complex work, break the task into ordered steps.
  • Use examples where they clarify the expected behavior.
Vague

“Summarize this document.”

Specific

“Summarize in exactly 3 bullets, each under 15 words, focused on financial risks. Output as a markdown list.”

Agent Instructions slide comparing vague and specific prompts
Craft

Behavioral Constraints

Constraints define the boundaries the agent must not cross, regardless of what a user asks.

  • Constrain scope, safety, sensitive data, actions, tone, and escalation.
  • Prompt-level rules alone are not enough; layer validation and guardrail models.
  • Put irreversible or costly action controls in code, not only in prompt text.
  • Red-team constraints and treat a bypass as a defect.
Behavioral Constraints slide describing agent guardrails
Craft

Persona Design

Persona is the voice and character users experience. It can improve consistency and brand fit, but it should never substitute for actual policy or controls.

  • Specify name and role, tone and voice, style, values, and boundaries.
  • Avoid over-acting a character when it reduces clarity or accuracy.
  • Use a reusable persona card so the specification stays consistent.
Example persona card
NameAriaRoleCustomer-support guideToneWarm, calm, conciseStylePlain language, no jargonNeverPromises refunds or invents policy
Persona Design slide with a reusable persona card
Prompt-Ops

Make prompts versioned, testable infrastructure

Prompt craft determines the behavior you want. Prompt-Ops makes that behavior reproducible, promotable, observable, and reversible.

Prompt-Ops

Prompt Versioning

A prompt version captures more than words. It should reproduce the conditions that produced the behavior.

Author & edit
Commit immutable version
Staging eval gate
Production
  • Capture prompt text together with model, parameters, tools, and metadata.
  • Use immutable IDs and clear diffs so behavior is reproducible and debuggable.
  • Promote through environments with quality gates and make rollback immediate.
Tools named in the deck: LangSmith · Langfuse · PromptLayer · Braintrust · Promptfoo · Maxim AI
Prompt Versioning slide showing author, commit, staging, and production flow
Prompt-Ops

Prompt Testing

Agent outputs are non-deterministic, so testing is based on datasets, scorers, baselines, and regression checks rather than exact-string assertions.

Dataset cases
Run variants
Score outputs
Compare to baseline
Ship or iterate

What makes a good eval

  • Typical queries, edge cases, and adversarial inputs.
  • Deterministic checks for structure plus model-based scoring for quality.
  • A baseline that new versions must beat.
  • Regression testing so a fix does not quietly break something else.

Close the loop

  • Score samples of production traffic with the same logic.
  • Turn low-scoring cases into new eval cases.
  • Grow coverage so known failures do not resurface.
  • Prefer grounded scorers over subjective impressions.
Prompt Testing slide showing dataset, variants, scoring, baseline, and release flow
Putting it together

The Prompt-Ops Lifecycle

Author, test, version, deploy, monitor, and improve. Production signals feed the next round of authoring.

Author
Test
Version
Deploy
Monitor
Improve
↻ Every production signal feeds the next round of authoring
Craft sets the ceilingA clear system prompt, precise instructions, firm constraints, and a consistent persona determine how good the agent can be.
Ops keeps it thereVersioning and testing keep quality from silently regressing as prompts, models, and data change.
Original deck

All 9 slides

Each slide uses responsive image variants from Slide600, Slide700, Slide900, and Slide1000.

1. Agent Development & Prompt Engineering

Click image to enlarge
Agent Development & Prompt Engineering slide

2. From Prompt Craft to Prompt-Ops

Click image to enlarge
From Prompt Craft to Prompt-Ops slide

3. System Prompt Design

Click image to enlarge
System Prompt Design slide

4. Agent Instructions

Click image to enlarge
Agent Instructions slide

5. Behavioral Constraints

Click image to enlarge
Behavioral Constraints slide

6. Persona Design

Click image to enlarge
Persona Design slide

7. Prompt Versioning

Click image to enlarge
Prompt Versioning slide

8. Prompt Testing

Click image to enlarge
Prompt Testing slide

9. The Prompt-Ops Lifecycle

Click image to enlarge
The Prompt-Ops Lifecycle slide
Enlarged slide