Systems on Glass is an AI-automation side project of mine. Every episode is generated by Claude (Anthropic's AI, working through Claude Code in my own video studio): the script, the animations and 3D scenes, the edit, and the narration in an AI clone of my voice. I choose the topics, bring the stories from my own work, and review each episode before it goes up. Not affiliated with Anthropic, OpenAI, Google or GitHub.
Every video in this series is made by an AI agent, and it never gets a clever prompt. Before it touches anything, it reads a file I wrote: 118 lines of who I am, what good looks like, and what it must never do. That file is a spec, and it beats clever wording every time.
The short version
- Vague prompts make the model choose. "Make me a video about indexes" leaves length, tone, level and voice open, so the model picks them for you. Every constraint you add removes wrong answers.
- The models got literal. OpenAI says GPT-4.1 follows instructions "more closely and more literally" than its predecessors, and Anthropic says Claude's latest models are "trained for precise instruction following". They do what you say, not what you meant. And contradictions cost: GPT-5 spends reasoning trying to reconcile them.
- A spec has five parts. The goal and why; the inputs and where the truth comes from; the rules, do and don't; the output format; and how you'll know it's done.
- Specs are the new code. Sean Grove (OpenAI) calls code "a lossy projection" of intent, and the tools now start from the spec: GitHub's Spec Kit, Kiro's requirements, design and tasks files, and AGENTS.md.
- Test it on a colleague, then keep it in a file. If someone with no context would be confused, so will the model. A file gets versioned, reviewed and reused.
My studio's spec: CLAUDE.md, 118 lines · this episode's storyboard: 29 slides, each with what I say, what appears, and the word it appears on
The same rule, twice. Anthropic's prompting docs show why a rule needs its reason: "never use ellipses" works less well than "your response will be read aloud by a text-to-speech engine, so never use ellipses". Before I'd read that, my studio's file already said: avoid ... mid-sentence, because it makes my cloned voice say "um". The reason is what makes a rule generalise. My Daily Bible workflow has the other kind of line, the kind about inputs: "Never invent or paraphrase scripture." The verse comes word for word from a Bible API, and the model only writes the scenes around it.
Further reading
- Claude prompting best practices — Anthropic's guide: be explicit, give the reason, and the colleague test.
- GPT-4.1 prompting guide — OpenAI on models that follow instructions more literally.
- GPT-5 prompting guide — why contradictory instructions cost more now, with the clinic example.
- Gemini prompt design strategies — Google on constraints, response formats and few-shot examples.
- The New Code (Sean Grove) — the AI Engineer World's Fair talk on specs as the real source code.
- GitHub Spec Kit — spec-driven development: a spec, then a plan, then tasks.
- Kiro: Specs — requirements, design and tasks files written before the code.
- AGENTS.md — an open format for giving coding agents their instructions: a README for agents.
Next episode: What an LLM call costs, and why long answers are the bill. Previously: Indexes. All episodes →