Start AI Coding from Zero

AI coding means building software with a model that can read requirements, inspect part of a codebase, propose and write changes, run commands, interpret errors, and help review the result. It is not a button that creates a perfect application. The outcome still depends on requirements, repository context, design direction, assets, permissions, tests, and human review.

1. Separate the Four Layers

Terminal agents are not necessarily IDEs. A coding agent and an editor can work side by side. Choose each layer deliberately instead of treating the brand name as a complete workflow.

2. What Is an LLM?

A large language model predicts and generates sequences of tokens. For software work, it can process natural language, source code, logs, diffs, documentation, and sometimes images. It does not automatically understand your project. It can only use what is available in the current context plus capabilities learned during training.

Ask the model to identify sources of truth, contradictions, missing information, and verification commands before it edits an unfamiliar repository.

3. Tokens, Tokenizers, Context, and Context Windows

A tokenizer splits input into units called tokens. A token is not always a word. Code punctuation, whitespace, and fragments of words can count differently across models.

Context is everything available for the current task: prompts, system and tool instructions, files, search results, logs, diffs, conversation history, and images. The context window is the maximum amount the model can process at once. A larger window helps, but sending an entire repository can still add noise, cost, and latency.

large repository
-> find relevant files
-> add requirements, rules, and evidence
-> work inside a focused context
-> verify the output

4. Why Markdown Matters

Markdown is plain text with lightweight syntax for headings, lists, tables, links, quotes, and code blocks. Files usually end in .md. It works well for PRDs, plans, repository instructions, design contracts, asset inventories, and runbooks because both humans and agents can read it and Git can show precise diffs.

Common files include README.md, PRD.md, DESIGN.md, PLAN.md, AGENTS.md, CLAUDE.md, GEMINI.md, and ASSETS.md. Auto-discovery differs by tool, so mention important files in the first prompt or follow the tool's official convention.

If raw Markdown syntax is unfamiliar, open the file in FileMira Markdown Viewer. It shows original and rendered views side by side in the browser. FileMira states that processing stays local in the browser, but you should still remove secrets before opening any file in a tool.

5. Minimum Safe Setup

For a local project, prepare Git, the required runtime, one IDE, one coding agent, an extracted project folder, safe secret handling, and known lint, typecheck, test, and build commands.

Create a recoverable baseline:

git status
git add .
git commit -m "chore: baseline before AI coding"

Do not initialize Git again if the folder is already a repository. Never commit .env.local, production credentials, or real customer data.

6. The Beginner Workflow

Use one small loop:

READ -> PLAN -> IMPLEMENT -> CHECK -> REVIEW DIFF -> COMMIT

Read the PRD, repository rules, schema, relevant files, and evidence. For a non-trivial change, agree on scope, exclusions, target files, risks, acceptance criteria, and commands. Implement one vertical slice. Run the commands defined by the repository. Then inspect git status and git diff before making a small checkpoint commit.

7. Start with Minimum Permissions

Allow repository reading and scoped working-tree edits. Review risky shell commands. Do not provide production secrets or production database access. Do not auto-approve destructive migrations, pushes, or deployments. A permission prompt is a chance to understand an action, not an obstacle to click through.

8. Design Still Needs Direction

A PRD explains what must work; it is not a complete art direction. Before generating many screens, collect two or three legal visual references, write DESIGN.md, identify production assets, implement one vertical slice, and perform human visual review. Better model capability does not replace typography, hierarchy, responsive behavior, brand assets, or accessibility decisions.

9. Ready-to-Start Checklist

• Requirements are available and conflicts are resolved.
• The repository has a clean Git baseline.
• Secrets are ignored.
• The agent knows the source of truth and task scope.
• Acceptance criteria and checks are explicit.
• UI work has references, a design contract, and an asset policy.
• The model or mode matches the task's difficulty and budget.

10. Continue the Path

Continue with Tools and IDE Setup, then Models and Limits, Markdown and Worktrees, Frontend Design, and Blueprint ZIP to Code.