Prompt Engineering

How to Write Agentic Prompts for DeepSeek-V4

The Clear AI Team June 28, 2026 8 min read

DeepSeek-V4 rewards structure. When you give it a role, a goal, and a set of tools, it plans before it answers. When you give it a vague question, it chats. This guide shows the exact scaffold we use internally at Clear AI to consistently coax agentic behavior from V4-class models.

The four blocks of an agentic prompt

Every reliable agentic prompt we ship is built from the same four blocks, in this order:

ROLE — who the model is playing
GOAL — the single outcome we want
CONSTRAINTS — the hard limits (time, cost, tools)
OUTPUT — the exact shape of the reply

Minimal template

textclear-ai

ROLE: You are a senior data engineer.
GOAL: Design a nightly ETL that syncs Postgres -> BigQuery.
CONSTRAINTS:
  - Budget: <$50/mo
  - Tools available: dbt, Airflow, Fivetran
  - Must handle 30M rows/day
OUTPUT:
  1. Architecture diagram (ascii)
  2. Cost breakdown
  3. Risks + mitigations

Why this works

Agentic models score every token against the plan they've committed to. If the plan is empty, the model fills it with pleasantries. Give it a plan and it fills it with execution.

“Structure is the shortcut to reasoning. Every constraint you add removes a branch the model has to guess.”
— Internal Clear AI eval, June 2026

Common failure modes

1. Overloading the ROLE block

One role per prompt. A model told it is simultaneously a lawyer, a poet, and a Kubernetes admin will hedge on all three.

2. Skipping OUTPUT

Without an explicit output shape, DeepSeek-V4 defaults to prose. Explicit outputs unlock JSON, code, tables — whatever downstream systems actually consume.

Tool-use tip: name the function, not the intent

Instead of saying 'search the web for pricing', say 'call web_search(query: str) with the vendor name'. Named calls trigger V4's tool-planning path.

pythonclear-ai

# Bad — intent only
prompt = "Find current pricing for Neon Postgres."

# Good — names the tool contract
prompt = """
Call web_search(query: str) exactly once with 'Neon Postgres pricing'.
Return the cheapest paid tier as JSON: {tier, price_usd, url}.
"""

Wrapping up

Agentic prompting isn't a secret incantation. It's the same discipline you'd use writing a ticket for a junior engineer: give them a role, a goal, the tools, and the output. Do that and DeepSeek-V4 will meet you where you are.