Case study: Kanban board, guided prompt + multi-model auto-fix

Same task - a single-file HTML Kanban board with working drag-and-drop, task add/edit/delete, and localStorage - sent to the 3 best-performing paid models from an earlier batch (mimo-v2.5, minimax-m3, step-3.7-flash), each once with a plain one-line instruction and once with guidance synthesized from a 3-LLM survey on prompting weak models. Every build was then actually tested in a real headless browser - click Add Task, type a title, check it really appears - not just read by another model. When the check failed, a fallback chain of models (mimo → minimax → laguna-xs-2.1:free → deepseek-v4-pro) tried to fix it, stopping at the first one that actually worked.

33%

Would've finished working, without the auto-fix loop

100%

Finished working, with the auto-fix loop

Builds that silently shipped a broken Add Task button on the first try

Biggest finding

Every one of the 4 broken builds got caught by the real browser check and fixed - none needed to go past the 2nd model in the fallback chain. mimo-v2.5 fixed 2, minimax-m3 fixed 2, and laguna-xs-2.1:free / deepseek-v4-pro were never even needed. A single text-reading reviewer model tried this same job in an earlier batch and fixed 0 of 3 broken builds - the difference here is testing the actual button click in a real browser, not asking another LLM to guess whether the code looks right.

Avg time, guided vs. plain

3.5m vs 3.2m

Excluding minimax's timeout outlier: 1.4m vs 3.2m - guided 55% faster

Avg cost, guided vs. plain

$0.018 vs $0.029

Excluding the same outlier: $0.014 vs $0.029 - guided 52% cheaper

Plain prompt (B)

Build a polished single-file HTML Kanban board with Todo, In Progress, and Done columns. Support drag-and-drop, task creation, editing, deletion, localStorage persistence, dark mode, and responsive design. Also show a small live task-count badge on each column header that updates immediately when tasks are added, moved, or deleted.

Guided prompt (A)

You are an experienced senior frontend engineer. Your job is to build production-ready code that compiles without placeholders.

Task: Build a polished single-file HTML Kanban board with Todo, In Progress, and Done columns. Support drag-and-drop, task creation, editing, deletion, localStorage persistence, dark mode, and responsive design. Also show a small live task-count badge on each column header that updates immediately when tasks are added, moved, or deleted.

Before writing code, briefly plan: what are the core interactions (add/edit/delete/drag) and what state needs to update together when each one happens (the board data, localStorage, and the visible task-count badges)?

Your task is complete only when:
- the file is created and is valid, complete HTML
- every interaction actually works when clicked/dragged, not just visually present
- no TODO comments or placeholder logic remain
- adding a task via the Add Task control actually results in a visible task on the board

Do not simplify the task or leave anything as "an exercise". Output only the plan and the code, nothing else.