AI Stack Builder

The Rules

Each job has its own ladder of plain rules, read from the top: if you already pay for a tool that does it, use that; if you asked for privacy, a tool that keeps your work private; if less than a set amount is left for this job, a cheaper or free plan; otherwise the tool built for it. Jobs run in your priority order, the first with all the money and each next with what is left. Every pick shows the step that chose it under "Why this pick", and "Why this stack" lists every decision in order. The whole rule list is public: stack-rules.json, with the tool list beside it, under CC BY 4.0. Prices come from the tool list, never from the rules.

When one step names more than one tool, the tie goes to the best score:

score = coverage × budget fit × privacy fit × (0.8 + 0.2 × signal)

• Coverage adds, for each thing you want done, how much better this tool does it than the stack so far (each tool is rated 1 to 3 per job), weighted by your priority: the first counts as many times as you ticked things, the last counts once.
• Budget fit is 1 for a free plan and falls toward 0.5 as the plan's price nears your whole budget. A plan that does not fit what is left is skipped, and the tool's free plan is used if it has one.
• Privacy fit is 1, or 0.6 for a cloud-only tool when you prefer tools that run on your machine.
• Signal comes only from public counts: the App Store rating weighted by how many people rated it, GitHub stars, and Hacker News stories in the last 30 days. Hacker News leans to developer tools, so a mention can lift the signal and never lowers it. Signal can move a score by a fifth at most, and a tool with no public counts gets 0.4 out of 1, a little under a tool with modest counts, since no counts is no evidence.

Once every job has its step, the builder spends the rest of your budget on the job you ranked first: it raises the plan of the tool that serves that job one tier at a time (when that tool has no bigger plan, the next job's tool instead), only to a plan whose extra it can say in one line, shown on the card as "Why this plan"; every other tool stays on the plan a small team starts on. How far it raises follows your answer to how much of your week the first job takes: a few hours keeps the starting plan, most days allows one tier up, all day goes as far as the money allows. Before each raise it compares, at that same price, every other tool built for the job and swaps to one that does better. A raised plan that already covers another job you ticked, such as ChatGPT Pro with Codex, replaces that job's pick, and one subscription is never paid twice. A pick that costs more than twice a cheaper tool built for the same job needs a reason it does more, or the cheaper tool takes the job; with privacy asked for, a free plan that makes your work public is passed over. A pick that no longer does any job best once the stack is in is dropped. When the first job's tool did not move up (no bigger plan, none that fits, or a plan you already pay for), it adds a second tool built for that job on its starting plan, but only for work like writing, coding, images or video, never a second store, domain, site, analytics, email or scheduler. When money is still left, it names the cheapest step up that fits, priced as the extra per month, or says nothing on the list would help. A tool bought once, such as Reaper, shows its price as paid once, outside the monthly total, and is offered only when a month of your budget covers it. The model line prices a rough monthly volume of tokens for your jobs at OpenRouter's live prices and picks, among four models, the one rated best for your kind of work (coding, writing or analysis) in LMArena's public categories whose cost stays under a quarter of what is left; its card names the rating, the rank and the leaderboard date. The model line appears only when you ticked writing copy, coding or building an app, and it is only for calling a model from your own code or automations, so it is always marked optional and never counted in the total; at a $0 budget it is not shown. There is no input for commissions anywhere in the formula. The scoring guide walks through each part.

Each pick opens a card with its public counts, marked live or dated, the latest Hacker News headlines about it, and notes on what people say. The notes are filed by hand from Hacker News comments, as a complaint or a highlight, each with its date, a link and the account that wrote it; the account was at least a year old when it posted and has no affiliate links in its profile, and the quote is under fifteen words. Notes never enter the score.

Related