{"slug": "are-you-controlling-your-llm-or-limiting-it", "title": "Are you controlling your LLM or limiting it?", "summary": "Developer Emile Silvis compared two approaches to building a Plants vs Zombies clone with LLMs: a highly controlled agent that grilled him on specs versus an autonomous agent given only a creative goal. The autonomous version produced a superior result in 25 minutes, leading Silvis to propose Reverse Engineer Driven Development (REDD), where a high-autonomy agent generates a prototype and a high-control agent rebuilds it carefully.", "body_md": "[Emile Silvis](/)\n\n# Are you controlling your LLM or limiting it?\n\nI had a realisation: being in control of your coding LLM is crucial, but you need to be clear headed about how much to control it, because you *are* limiting it in ways that you wouldn't expect.\n\nThis weekend, I wanted to learn more about [Matt Pocock's /grill-me-with-docs skill](https://www.aihero.dev/grill-with-docs). Before any code exists, the agent interviews you relentlessly about your plan, one question at a time, and writes the answers down as it goes. You end up with a glossary, and hard-to-reverse decisions land as architecture decision records. Any future agent runs can reference these artifacts.\n\nTo test it out, I chose a semi-ambitious project (not a straightforward web-based CRUD): making a clone of Plants vs Zombies.\n\n~4 hours later, I ended up with this:\n\nAt this point I got curious: what could a powerful model do if I gave it *no* constraints? So I gave Claude Code Fable a single `/goal`\n\n.\n\n```\n/goal a cute-as-heck pixel-based Plants vs Zombies game where frogs protect their homes (mushrooms) from monsters. Fireflies are the currency (sun in PvZ). Ambient soundcape with crickets and other nature sounds, generative. Cute as heck. Sounds are cute as heck too. Game should be super duper cozy and beautiful. Runs in a browser.\n```\n\nIn ~25 minutes, it came up with this (it's a video — press play!):\n\nThe `/goal`\n\nversion was better in almost every conceivable way. Smoother, more polished, cuter. It even had a cool soundtrack!\n\nMy expectation was that the `/grill-me-with-docs`\n\nversion would *eventually get there*, but once I saw the `/goal`\n\nversion, it became clear that it never would: the `/grill-me-with-docs`\n\nversion focussed so hard on not straying from the spec, that it developed a tunnel vision and stifled its own ability to produce code that can even potentially be iterated into what the `/goal`\n\nversion came up with.\n\nSo what's the lesson? There's a spectrum, and you need to be clear headed about where you want to be, and why.\n\nWhen you have this spectrum in mind, you can choose the appropriate strategy:\n\n- If it's mission-critical software, err on the side of control.\n- Creative green fields, go for autonomy.\n\nBut what's most interesting to me is doing both: describe your dot on the horizon and let a high-autonomy mode agent build out an all-bells-and-whistles prototype. Then ask a high-control mode agent to lock that prototype in as its target, reverse engineer it, and start carefully from scratch. You first generate the vision, then build carefully towards it. Call it **Reverse Engineer Driven Development (REDD)**. I'll definitely be exploring this approach more in the future.", "url": "https://wpnews.pro/news/are-you-controlling-your-llm-or-limiting-it", "canonical_source": "https://emilesilvis.com/are-you-controlling-your-llm-or-limiting-it.html", "published_at": "2026-07-08 00:00:00+00:00", "updated_at": "2026-07-08 10:14:34.179846+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-tools", "developer-tools"], "entities": ["Emile Silvis", "Matt Pocock", "Claude Code Fable", "Plants vs Zombies"], "alternates": {"html": "https://wpnews.pro/news/are-you-controlling-your-llm-or-limiting-it", "markdown": "https://wpnews.pro/news/are-you-controlling-your-llm-or-limiting-it.md", "text": "https://wpnews.pro/news/are-you-controlling-your-llm-or-limiting-it.txt", "jsonld": "https://wpnews.pro/news/are-you-controlling-your-llm-or-limiting-it.jsonld"}}