Vibe Coding 101
The first lesson for newcomers: let the agent write code from natural language
"Vibe Coding" means describing the outcome you want in natural language and letting the AI handle the actual code and file operations. DeepSeek Harness is built for exactly that: the agent reads files, edits files, runs commands and plans its own work. This tutorial takes you through your first complete Vibe Coding experience.
This tutorial assumes you have finished installation and model setup, and have selected a workspace.
Step 1: Pick a small, concrete task
Don't start with a big goal. Good starting points:
- "Add an installation section to this README"
- "Explain what each directory under
src/does" - "Extract the duplicated code in this script into a function"
The task should have a clear definition of done. A vague "optimize my project" makes the agent guess repeatedly.
Step 2: Give the agent context
The agent's view is the workspace plus your description. Include key information in the task:
Describe first, then expect
Don't write just "fix the bug". Write: "Running pnpm test fails with an error in src/parser.ts; locate the cause, fix it, and make the tests pass."
Step 3: Understand approvals
Before the agent performs sensitive actions like writing files or running commands, the Web UI raises an approval request. Principles:
- Review the diff and the command before approving;
- If you don't understand it, ask the agent to explain before executing;
- Reject uncertain write actions and ask for analysis only.
Step 4: Iterate in small steps
- Change one thing at a time and verify immediately;
- Let the agent run tests or builds to verify its own results;
- Keep conversing if unsatisfied — the agent corrects course from the current state.
Step 5: Use skills and goals
- Skills: make the agent follow domain best practices; for example the
code-reviewskill reviews changes by review standards; - Goals: give one long-running objective and the agent advances it across rounds automatically — see skills & goals.
Safety advice
- Practice on a separate branch or a copy of the repository;
- Back up or commit important files first;
- Approvals are your last line of defense — never blindly "allow all".
Next: AI office work, or the more advanced CLI & profiles.