Karmaşık görevlerde zincirleme prompt'lar nasıl etkili olur?

15 Eyl 07:54

I tried exactly the three‑stage workflow you described while building a code‑generation assistant for a Flutter project. First I fed the model a concise “what‑to‑build” spec, then I asked it to produce a short bullet‑point summary of the required widgets, and finally I prompted it to expand each bullet into full Dart snippets with comments. The trick that kept the context from drifting was to carry over a **compact summary** (the second step) as a “memory token” that I prepend to every subsequent prompt, rather than re‑sending the whole original spec. This way the model always sees the high‑level intent and the current focus without blowing the token budget. To further minimise context loss, I use a deterministic identifier for each step (e.g., `STEP1:`, `STEP2:`) and store the model’s output in a small JSON block that I parse and re‑inject. Adding a brief “recap” line like “Remember we need a responsive list view from step 2” before the next prompt helps keep the narrative consistent. In practice, this staged approach uses far fewer tokens than a single massive prompt and gives you more control over where the model might hallucinate. So yes, for complex tasks with clear sub‑goals, chaining prompts tends to be more efficient and reliable than cramming everything into one long request.