A month ago I found myself half-leaning, half-lounging at my desk, nursing a tepid latte and wondering whether terminal phobia is the last socially acceptable tech anxiety. So I tried an experiment:
Can a self-described “sufficient amateur” ship a public-facing site—design, deploy, everything—using nothing but Cursor + Claude Code?
The result now lives at academicinnovation.support (this website). No formal front-end and design training. Just decent taste from magzine reading and a willingness to let Claude models wander down dead ends while I watched. Cursor’s interface made the whole thing feel like pair-programming with a caffeine-addled librarian who never forgets a footnote.
The Art of Reflective Prompting 🎨
You might think productive coding sessions ended the moment the script worked. Agentic coding flips that script:
- Celebrate the success.
- Immediately interrogate the tool about what it didn’t do.
In my example-download repo—a little summer-school side quest to batch-fetch files—the post-victory prompts were where the real tuition happened:
“Claude, what feature of Claude Code would have saved us time here?”
“Show me, step-by-step, how to /fork, explore a crazy refactor, then rewind without ruining main.”
Claude answered with mini-tutorials on /plan, CLAUDE.md, and, yes, the mysterious /fork command. Those chat snippets crystallised into CLAUDE_CODE_BEST_PRACTICES.md, a living guide that sits beside the code and evolves exactly as the project (and my understanding) evolves.
The meta-lesson: every finished script is a launchpad for questions that tailor the next lesson to you—not the syllabus, not the generic MOOC, you. That is the quiet super-power of agentic workflows: instant feedback loops that write their own curriculum.
Terminal Ice-Breaker 🤖
At first glance image-cli is laughably simple: a Bash wrapper that flings a PNG up to Wolfram Cloud and opens the resulting URL. Yet the README reads like a mini-textbook on Unix fundamentals: shebangs, $PATH, symlinks, regex, sudo, quick-n-dirty Git hygiene.
For a humanities student who thinks “ls” is a New York subway line, running this script is the moment: terminal horror melts into muscle memory. And because Claude is lurking, the learner can ask:
“What exactly lives inside /usr/local/bin and why should I care?”
Claude obliges with comparisons, diagrams, even PowerShell equivalents—adapting the micro-course to whatever background knowledge (or lack thereof) the learner brings.
Best Practices 🙊
- Begin with
/plan. Think of it as the syllabus Claude drafts before you open a code file. - Drop a
CLAUDE.mdin the repo early. It becomes the shared long-term memory—like leaving sticky notes for your future self. - Reflect out loud. After each milestone, ask Claude to critique the workflow, grade its own output, or propose a five-minute classroom demo. Those reflections are the learning artefacts.
Why Being a “Sufficient Amateur” 🐣
That patchwork curiosity is the soil where agentic coding thrives. Claude Code doesn’t reward rote memorisation of APIs; it rewards good questions. Ask them, and the AI will happily split into ten conversational branches, each tailored to your next aha-moment.
Which means a trivial script—two dozen lines tops—can morph into a bespoke micro-course on shell philosophy, Git etiquette, even deployment strategy. The cost? A cup of coffee and the nerve to type “/fork”.
So uncap your latte, open Cursor, and dare Claude to teach you something it forgot to mention in the docs. Your future self (and your terrified first-time terminal students) will thank you.