Agentic AI: Prompt Engineering & Context Engineering

1 Core definitions: prompt vs. context

Prompt engineering is the craft (and increasingly the discipline) of writing instructions that steer model behavior toward a desired outcome—clarity, constraints, examples, and output formats. Think of it as the interface design between a user (or system) and the model.

Context engineering is the system-level discipline of ensuring the model is given what it needs to plausibly accomplish the task: relevant knowledge, policies, tool schemas, state, memory, and constraints within a token budget. It’s less “wordsmithing” and more “systems engineering of the context window.”

Context Engineering Visual Prompt vs. Context Engineering (img1.png)

Context

2 What “Agentic AI” changes

Agentic AI typically means a workflow that makes multiple LLM calls and/or uses tools (search, databases, code execution, ticketing systems, etc.) to achieve a goal—often with planning, iteration, and verification.

In a single-turn chatbot, the prompt dominates. In an agentic workflow, three things become equally important:

• State: what we already know (user preferences, prior steps, partial results).
• Capabilities: tool access + tool contracts (what functions exist and how to call them).
• Control: guardrails, stop conditions, evaluation checks, and “what good looks like.”

GOAL → PLAN → (RETRIEVE CONTEXT) → ACT (tool calls / drafts) → VERIFY → REFINE → DONE
        ↑______________________________________________________________↓
                 If verification fails, update context + retry

Prompt engineering is how you communicate, while context engineering is how you engineer the communication environment so the model can succeed.

Context Engineering Visual RAG vs. Context Engineering (img2.png)

Context

Context Engineering Visual Context Engineering diagram (img3.png)

Context

3 How they are usually designed (a practical blueprint)

High-performing agentic systems are built from a small set of reusable design choices. A helpful way to think about it is: you are designing artifacts and a lifecycle.

A) Prompt engineering artifacts (agent-ready)

• System charter: role, boundaries, safety constraints, refusal rules, tone.
• Task spec template: objective, inputs, outputs, acceptance criteria.
• Output contracts: JSON schema, Markdown sections, “must include / must not include.”
• Few-shot exemplars: minimal, representative, and aligned with the output contract.

B) Context engineering artifacts (production-grade)

• Retrieval pack: curated sources, embeddings, ranking, filtering, citations/provenance.
• Tool pack: function definitions, constraints, examples, error-handling expectations.
• Memory pack: stable user facts, project state, summaries (with freshness rules).
• Token budget: reserved space for instructions, state, retrieved evidence, output.

4 Examples you can reuse

Below are agentic-ready templates that show the difference between “prompting” and “engineering the context the agent needs.”

Example 1 — Prompt template for an agentic task

SYSTEM (stable):
You are an execution-focused AI agent. Follow policies. If missing data blocks correctness, ask for it.
Prefer short, structured outputs. Use tools when needed.

DEVELOPER (stable):
Output must match the requested schema. Cite sources when you use retrieved evidence.
Never invent identifiers, dates, or quotes.

USER / TASK (variable):
Goal: {one sentence objective}
Context: {what the user already decided + constraints}
Inputs: {files, links, data, system state}
Definition of Done:
- {acceptance criterion 1}
- {acceptance criterion 2}
Output format:
- {sections or JSON schema}
Quality checks:
- {e.g., "verify totals", "include assumptions"} 

Example 2 — Context pack for a “budgeting assistant” agent

CONTEXT PACK (assembled at runtime)
1) Policies
   - Allowed advice scope (finance/legal disclaimers)
   - Safety constraints
2) State
   - User locale, currency, current month, known accounts
   - Last computed budget and assumptions
3) Retrieval (RAG)
   - Relevant policy docs, prior notes, verified rates
   - Each snippet has: source_id, timestamp, excerpt
4) Tools
   - get_transactions(account_id, date_range)
   - categorize_transactions(list)
   - forecast_cashflow(history, horizon)
   - Each tool includes schema + examples + error cases
5) Token budget plan
   - Instructions: 10%
   - State: 15%
   - Retrieval evidence: 35%
   - Working space: 10%
   - Final answer: 30%

Example 3 — Agentic “verify then finalize” pattern

This is a common pattern for making agents more dependable: the model produces a draft, then runs a verification step with explicit checks (often deterministic), and only then finalizes.

PASS 1 (Draft):
- Produce output quickly using current context.
- Mark uncertainties explicitly.

PASS 2 (Verify):
- Run checklist:
  • Are required fields present?
  • Are numbers internally consistent?
  • Are claims supported by provided evidence?
  • Did we respect constraints (tone, length, exclusions)?
- If any check fails: retrieve missing context OR ask a targeted question.
- Then produce final output.

5 A quick production checklist

If you adopt only one habit: treat your context as a first-class engineering artifact—version it, test it, and keep it lean.

5 A quick production checklist

If you adopt only one habit: treat your context as a first-class engineering artifact—version it, test it, and keep it lean.