I'm not an expert by any means, but I feel like this has worked for me. First off my core system prompt (or similar) contains a piece about how I want it to be critical of what I say or suggest, and to not assume what I say is a good idea.
Unless I have a very clear plan for what I want to build, I say I want to brainstorm X. I'll say what I had in mind, ask what it thinks. We then go back and forth, I might ask for alternatives and such along the way. Once I know what I want to build, I say make a detailed design document of what we've arrived at, include rationales for the decisions made. If it's a larger project I say to split the design into phases that can be implemented incrementally. The document is written to disk.
I'll look over the design document and perhaps do a few rounds of tweaking on it.
Once I'm happy I switch to plan mode, possibly with a fresh context, and say to make a detailed implementation plan for phase 0. I'll look it over and adjust until I am happy. If it didn't already I'll have it split the work into concrete steps or deliverables. This allows me to keep tabs on what it's doing and intervene if it goes off the reservation.
I'll tell it to write the implementation plan to a file before it starts work. After it's done if it didn't automatically I'll have it write a summary of the work done and issues that came up.
These summaries along with the main designs document becomes input to subsequent phases.
This has worked well for me with both Codex and GPT 5.x and OpenCode with Qwen 3.6 27B.
I've been telling Gemini to limit it's response to a certain amount of words, typically in 25 word increments. When I just want it to agree with me, I add "25 words." to the end of the prompt. When I want it to give me a rough answer, 50 - 75 words. For a more robust response, 100 - 200 words. This has made my experience much, much better.
That is a great pattern. Word count constraints force the model to prioritize instead of padding. I use a similar idea but in the other direction -- asking for the single most important thing rather than a list, so the model has to commit to one answer rather than hedge across five.
I'm not an expert by any means, but I feel like this has worked for me. First off my core system prompt (or similar) contains a piece about how I want it to be critical of what I say or suggest, and to not assume what I say is a good idea.
Unless I have a very clear plan for what I want to build, I say I want to brainstorm X. I'll say what I had in mind, ask what it thinks. We then go back and forth, I might ask for alternatives and such along the way. Once I know what I want to build, I say make a detailed design document of what we've arrived at, include rationales for the decisions made. If it's a larger project I say to split the design into phases that can be implemented incrementally. The document is written to disk.
I'll look over the design document and perhaps do a few rounds of tweaking on it.
Once I'm happy I switch to plan mode, possibly with a fresh context, and say to make a detailed implementation plan for phase 0. I'll look it over and adjust until I am happy. If it didn't already I'll have it split the work into concrete steps or deliverables. This allows me to keep tabs on what it's doing and intervene if it goes off the reservation.
I'll tell it to write the implementation plan to a file before it starts work. After it's done if it didn't automatically I'll have it write a summary of the work done and issues that came up.
These summaries along with the main designs document becomes input to subsequent phases.
This has worked well for me with both Codex and GPT 5.x and OpenCode with Qwen 3.6 27B.
I've been telling Gemini to limit it's response to a certain amount of words, typically in 25 word increments. When I just want it to agree with me, I add "25 words." to the end of the prompt. When I want it to give me a rough answer, 50 - 75 words. For a more robust response, 100 - 200 words. This has made my experience much, much better.
That is a great pattern. Word count constraints force the model to prioritize instead of padding. I use a similar idea but in the other direction -- asking for the single most important thing rather than a list, so the model has to commit to one answer rather than hedge across five.
The useful prompts are usually boring. Give context, say what you want, say what to avoid, and set the tone. Without that it often gets too polished.
[dead]
[flagged]