هندسة البرومبت المتقدمة مع Claude: التفكيك والتثبيت وسلاسل التفكيرAdvanced Claude Prompt Engineering: Decomposition, Anchoring and Chain of Thought

Key terms

Decomposition

Asking Claude to split a big task into ordered steps first, then running each step with its own prompt.

Anchor statement

One short instruction that fixes three things: the role Claude plays, the main constraint, and the output format.

Drift

In long conversations the model slowly moves away from your original instructions.

Chain of thought

Making Claude write out its reasoning steps before the final answer, so you can check and correct each step.

System prompt

Instructions Claude reads before your messages. In Claude Code you can pass one with a flag. In the chat, project instructions play the same role.

The steps

1. 1

   Step 1

   Split the task before you ask

   Claude does its best work on one specific task at a time. Pack ten requirements into one prompt and every part comes back average.

   Decomposition fixes this. First ask Claude for a plan only, then run each step with its own prompt. Claude tracks one objective at a time, and you can review and correct each output before moving on.

   Paste the prompt below in a new chat, fill the task, and read the plan before you run anything.

   PromptDecomposition plan

   Copy

   You are a senior software architect. Decompose the following task into discrete, ordered subtasks.
   
   Task: [paste your task here]
   
   For each subtask give:
   1. One clear objective (one sentence)
   2. Required inputs (information or files needed)
   3. Expected output (what gets produced)
   4. Dependencies (which earlier subtask must be done first)
   
   Do not start executing any subtask. Only produce the plan. Number the subtasks and keep the total between 4 and 7. If a subtask needs a human decision or outside input, mark it [HUMAN CHECKPOINT].

   Tip: If a task takes more than two paragraphs to describe, that is the signal to split it.