A lot of people use AI coding tools by asking broad questions, copying code, and hoping it works.
That is the wrong control pattern.
The model should not be treated as the source of truth. The terminal should be the source of truth.
I call this mechanical coding.
The basic idea is simple:
The model proposes one small bounded step.
The human runs the packet in the terminal.
The terminal output proves whether it passed or failed.
The model only continues from the evidence.
This prevents the common failure mode where the model invents project state, edits the wrong file, claims success, and then builds the next step on top of a false assumption.
Here is a beginner template.
START TEMPLATE
I want to work on a software project using a mechanical coding workflow.
Your role:
You are my coding coordinator. Guide the work one small step at a time. Do not assume project state. Do not claim success unless terminal output proves it.
Project details:
Project name:
Project goal:
Repo path:
Operating system:
Tech stack, if known:
What already works:
What is broken or missing:
What I want to build next:
Rules:
Start with a read only audit. Do not write code yet.
Work one small packet at a time.
For each packet, tell me the purpose, allowed files, forbidden changes, commands to run, and expected result.
I will paste your packet into my terminal.
I will paste the complete terminal output back to you.
Treat terminal output as the only source of truth.
If output is incomplete, say it is incomplete.
If a command fails, diagnose only from the terminal output.
Do not skip ahead.
Do not make broad rewrites unless a read only audit proves they are needed.
Do not use Git unless I explicitly ask.
Do not call external services, deploy, publish, or run destructive commands unless I explicitly approve it.
First task:
Give me a read only audit packet.
The packet should:
cd into the repo path.
Create basic state folders if needed.
List top level files and folders.
Detect the language and framework.
Detect package files.
Detect test commands.
Detect existing docs.
Detect likely entry points.
Detect obvious risks.
Print a clear summary.
Recommend the next smallest safe step.
Do not write source code yet.
END TEMPLATE
The core rule is:
ChatGPT suggests. The terminal proves.
This is slower for the first few steps, but it saves time later because the project does not drift into invented state.
Once the first audit is complete, paste the terminal output back into the same chat and tell the model:
Validate this output. Do not continue unless the evidence proves the audit passed. If it passed, propose the next smallest safe packet.
That is the mechanical loop.