A loop that learns
Change one variable per cycle. Plan, do, check, act. A loop that learns beats a line that runs.
Your job: turn the weakest dimension into a single-variable experiment.
Help me iterate my page My page scores [X], weakest on [dimension]. Propose one — only one — variable to change, write the new version in my voice, and tell me which metric should move if the hypothesis holds. claude mcp add --transport http leanspark https://leanspark.ai/mcp Connect first →
What you get back
Takes the weakest dimension from your last score, turns it into a one-variable hypothesis, drafts just that change in your voice, and re-scores after — so each cycle teaches you exactly one thing.
A single-variable change, a before-and-after score, and a keep / kill / scale call. You ship the change and read the live metric.
A landing page is never finished; it’s iterated. The trap is iterating without learning. If you change the hero copy, the image, the problem framing, the proof, and the button color all at once, and the new version does better — you’ve learned exactly nothing. You can’t tell which change moved it.
One variable per cycle. Change one thing. Measure. Then change the next. It feels slow and it’s the fastest way to actually learn what moves your page, because every result points at a single cause.
The loop is PDCA — Plan, Do, Check, Act — the same learning cycle Deming carried to Japan in 1950 and Toyota built its production system on. On a landing page it reads cleanly:
- Plan — form a hypothesis specific enough to be wrong. “If I name the segment instead of saying ‘founders,’ Targeting moves from 7 to 9.” The brief is the plan.
- Do — build the smallest version of that one change. Not the whole page.
- Check — score it, and read the metric you predicted would move. Honestly.
- Act — keep what worked, discard what didn’t, update the brief, run the next cycle.
After a few cycles you’ll have a verdict: kill it, keep iterating, or scale it by turning up the traffic you held back until the page was ready.
The Check step now comes to you. In LEANSpark, a page ships inside a conversion experiment with a structured goal — a target rate the page has to hit, a visitor floor below which there’s no verdict, and a time window. Recording the page starts the clock. Every three days the system posts a check card: report your numbers. Your AI assistant pulls the real analytics and reports them back over the same MCP connection, and LEANSpark scores them against the goal — on goal, off track, or not enough traffic yet. The loop you used to run on willpower now runs on a schedule. One honesty rule is enforced: real counts for real periods. The system doesn’t take estimates.
The agent proposes. The founder decides. Your AI assistant can stage an experiment for you — template, hypothesis, goal — over MCP. What it can’t do is start one. A staged experiment sits in your sprint until you review it and press start, because the moment an agent can silently start and stop experiments, you’ve automated lying to yourself. Drafts and proposals are the agent’s job. Decisions stay yours.
Put the page on the clock. The second exercise of this chapter: stage a Landing Page Conversion Test in your current sprint — hypothesis, target rate, visitor floor, time window — and file your page into it. The experiment lands as a proposal: nothing runs until you review it and press start. Via MCP: “Stage a landing page conversion experiment for this page over MCP, with my hypothesis and a target rate worth beating — then file the page into it with record_artifact. Tell me what it returned, including who has to start it. When the check cards arrive, pull my real analytics and report them with report_metric — real counts for real periods only.”