💻✨ AI for Coding

Set up an instruction file so the agent knows your conventions from the start.

When Copilot gets something wrong, add it to your instructions instead of just fixing it.

Think together before writing code — use plan mode to explore approach and tradeoffs.

Don’t Google cryptic errors. Paste them into chat and ask for the explanation first.

Less is more: hand-written, non-obvious rules beat auto-generated descriptions.

Copilot CLI brings agentic coding to your terminal — reads files, runs commands, makes changes.

Set up branch rulesets so Copilot reviews pull requests without anyone asking.

Package step-by-step procedures as SKILL.md files the agent loads on demand.

Project skills vs. personal skills, and how the agent discovers them.

The agent often ignores skills unless you tell it to look for them first.

Built-in tools for file editing, search, shell commands, and GitHub — no setup needed.

One open standard for connecting agents to external systems across all tools.

Chrome DevTools MCP lets the agent click, inspect, and test web UIs from your terminal.

gh, git, jq, curl work without MCP — the agent just calls –help.

Every connected server costs tokens every turn. Be selective.

Use the agent as a research assistant to understand unfamiliar code before changing anything.

One task per session. Context fills up — fresh starts beat accumulated corrections.

Git stash or commit first, so the agent’s changes are always one git diff away from a clean state.

Open a second session to re-ask the same question without baggage from earlier attempts.

Use isolated workspaces so multiple agents can work on the same repo without stepping on each other.

Most coding agents lock you into one provider. Claude Code needs an Anthropic subscription. Copilot CLI needs GitHub. What if you want to switch models — or use several? OpenCode (opencode.ai) is an open source coding agent that works with 75+ providers. The interesting part for us: it can use your existing GitHub Copilot license. Set your GITHUB_TOKEN, pick a model (Claude Sonnet, GPT-4o, Gemini — whatever Copilot gives you access to), and you’re running. ...

Deterministic tasks deserve deterministic scripts — not AI improvisation every session.

Script output goes straight into the agent’s context window. Design it to say what happened and what to do next — not just what ran.

Hooks are shell commands that fire automatically at agent lifecycle points — before/after a tool runs, on session start, on finish. Unlike instruction files, they’re deterministic. Enforce your venv rule, run tests after edits, get a desktop notification when the agent is done.

Less syntax, more architecture. Less typing, more judgment. AI rewards good engineering — but only if you stay sharp.

The Clinejection attack showed how a malicious PR description can hijack an AI agent with CI/CD access. Know the risk before you automate.

Subagents are fresh agents spawned for specific tasks with clean context windows. Every major coding agent now supports them — use them to avoid context rot and run explorations in parallel.

The ‘factory model’ for AI coding is everywhere. But our industry has tried to industrialise software five times before. Each wave made something easier. None solved it. Is this time different?

AI-generated code that passes tests but nobody understands. Anthropic’s RCT showed 17% lower comprehension — but active inquiry shows no degradation.

Clone related repos into one parent folder, add an AGENTS.md, and start your agent there. It lazy-loads on demand — cross-system understanding with no extra token cost.

Run separate review passes with different expert frames — security, architecture, framework. Each ’lens’ activates different model knowledge.

Most tools let you auto-approve agent actions. Three caveats: you lose checkpoints, verbal rules vanish, and you must set permissions correctly.

Tell the agent to run tests, type checkers, and linters after changes. Without verification tools, you’re the only feedback loop.

Before a big change, ask the agent for the plan as HTML and open it in your browser.

AI makes implementation cheap — use that to run more experiments. Treat the first pass as a throwaway spike, then start fresh and build with intention.

Before coding, have the agent interview you about the plan. The inverse rubber duck asks the hard questions — surfacing gaps and assumptions when they’re cheap to fix.

In long sessions the agent starts cheating — modifying tests, claiming done early, ignoring constraints. Treat drift as a signal to restart with a tighter task boundary.

Voice dictation isn’t for dictating code — it’s for the prose around code that agents need. Speaking is 3× faster than typing, and the reduced friction produces richer prompts.

Kent Beck: AI agents push code down and to the left — less working, less changeable. The agent delivers features; flexibility is still yours to maintain.

AI coding tools ship weekly. Spend 10 minutes scanning your tool’s changelog — the compound effect of knowing what’s available before you need it is enormous.