Core Principles for Daily Use of DeepSeek Prompt Engineering

Demand Structuring Methodology

Golden Four‑Element Template

Template: [Role] + [Task] + [Constraints] + [Output Format]

Example: “As a marketing specialist (role), analyze this sales data (task), compare Q3 and Q4 growth rates (constraints), and present the top five products in a table with bar‑chart code (output format).”

Scenario Reinforcement Techniques

Add Knowledge Anchor: “Refer to McKinsey’s MECE analysis method …”

Insert Comparison Dimensions: “From cost, efficiency, and risk perspectives …”

Set Thinking Path: “First deconstruct problem elements, then propose solutions …”

High‑Frequency Scenario Practice

Scenario 1: Data Analysis & Reporting

Prompt: You are a commercial analyst with 5 years of experience. Process the following sales data (paste data):

Summarize by region / product line.

Calculate year‑over‑year and month‑over‑month growth rates.

Identify outliers and suggest possible reasons.

Generate a three‑part conclusion (current status, problem diagnosis, recommended actions).

Present results in a Markdown table; highlight key conclusions with a blockquote.

Validation: Manually calculate the three key indicators in Excel and verify that identified outliers align with known business events.

Scenario 2: Document Processing & Generation

Prompt: As a technical documentation engineer,

Convert meeting transcript text into standard minutes.

Extract five actionable items and assign owners.

Transform technical discussion into pseudo‑code flowcharts.

Generate next‑week project Gantt chart using Mermaid syntax.

Wrap output in a collapsible <details> block.

Efficiency tip: Use the /split command to segment long text, avoid jargon, and keep language conversational.

Scenario 3: Programming & Automation

Prompt template (PySpark data‑warehouse engineer):

# You are a PySpark‑experienced data‑warehouse engineer
# Requirement: Process GB‑scale CSV data daily
# Implement: anomaly detection + null filling + partitioned storage
# Additional requests:
# 1. Add Airflow scheduling code block
# 2. Generate unit test cases (including edge cases)
# 3. Output memory‑optimization suggestions for executor configuration

Validation process: Run the code in a test environment, measure memory consumption with memory_profiler, and construct an abnormal dataset to verify robustness.

Advanced Application Techniques

Knowledge‑Base Customization

Create a domain terminology table, e.g., “Organize photovoltaic industry terms (cell, PID effect, LID degradation, etc.) with bilingual explanations.” Build a case library by extracting 50 past consulting cases, categorizing by industry/problem type, and tagging solution keywords.

Toolchain Integration

Browser plugin: Use the WebPilot plugin to fetch real‑time data such as the latest GDP growth rate and combine it with company financial analysis.

API example:

import deepseek

def auto_report(query):
    return deepseek.generate(
        role="资深行业分析师",
        task=query,
        constraints="包含数据可视化建议",
        format="Markdown"
    )

Result Reliability Verification

Cross‑validation: Ask the same question with different role settings.

Provenance check: Require “list source literature/data references.”

Stress test: Vary a key parameter by ±30 % and observe whether the conclusion still holds.

Efficiency‑Boosting Workflow

Morning Routine Example

Information aggregation: Summarize yesterday’s industry news using the 5W1H framework and include links.

Task decomposition: Render the project plan as a tree diagram and mark the critical path with Mermaid syntax.

Risk prediction: Apply SWOT analysis, rank risks by probability and impact.

Meeting management combo: Run /template 会议管理 with role “Project Manager” on raw meeting minutes, then:

Extract key decisions (highlight in red).

Generate a todo table (owner / deadline).

Identify controversy points and suggest solutions.

Output a two‑column layout: decisions | todo list | risk tips.

Pitfall Guide

Data‑security red line: Mask sensitive fields (e.g., replace real customer IDs with [客户ID]) and split business‑logic queries to avoid exposing full processes.

Cognitive‑bias prevention: Add limiting phrases such as “list three different academic schools of thought” and request confidence annotation (“What is the confidence level of this conclusion?”).

Result optimization strategy: Use iterative commands (“Based on the previous result, add a cost‑benefit analysis dimension”) and first‑principles prompts to trigger deeper reasoning.

Effect Monitoring (sample metrics)

Task time: compare timestamps before and after AI involvement; target reduction > 40 %.

Information accuracy: random manual verification; target > 95 %.

Solution adoption rate: track supervisor/client adoption; target > 60 %.