CRUX — CRUX: An AI advisor that does not make the user do the work

The Problem

Selecting a workflow automation platform takes hours of research and still produces the wrong answer

The workflow automation market has dozens of competing platforms, each with a different pricing structure, integration count, execution limit, and hosting model. A 5-step workflow running 1,000 times per month costs 5,000 tasks on Zapier and 1,000 executions on n8n. That structural difference alone produces a 10x cost variance for the same workload. Most buyers have no framework for making that comparison accurately.

The problem surfaced consistently while consulting for technology companies on automation strategy. Clients had a clear automation goal. The obstacle was platform selection, not the automation itself. No neutral, accurate, up-to-date resource existed to make that comparison efficiently.

General AI tools return pricing estimates based on training data, which is often months out of date. Comparison sites carry affiliate bias. CRUX was built to address both problems.

Screenshot of crux.do — the single input field and prompt chips — showing how the entry point is as simple as the problem is complex. — The landing page entry point at crux.do

Brand Identity

Name, mark, and visual system built on one precise concept

"CRUX" is defined as the decisive point of a matter. The brand was built on that definition. CRUX identifies the correct platform from a set of competing options. The name is the function.

The logomark was designed as a four-stage progression: scattered inputs, convergence, a unified form, and a star at the center representing the decision point. The mark is a visual diagram of the product's core process, not a decorative element.

The visual system uses dark backgrounds with a green primary for technical legibility. Nunito Sans provides readability at small sizes across both light and dark modes. The system had to function in a technical context without the aesthetic of a developer tool.

The logomark conception and color system from the brand guidelines

Interface Design

Reducing decision-making friction at every interaction point

Most AI-powered interfaces open with a blank input and no guidance. Users are expected to know how to ask the right question, at the right level of specificity, using the right vocabulary. For a platform comparison tool, that approach fails at the first interaction. Most users do not know which variables are relevant or how to frame a query that produces an accurate recommendation.

Context built through sequential, single-question dialogue

CRUX collects requirements one question at a time, each informed by the previous answer. Users are never presented with a form or multiple simultaneous questions. By the point of recommendation, the system has captured the user's stack, volume, budget, and technical level without the user making any structural decisions about how to frame their request.

Predictive prompt suggestions at every entry point

The landing page and conversation entry points surface pre-populated example queries derived from platform usage data and search trend analysis. These are not generic placeholders. They reflect the actual questions users search for most — cost comparisons at specific volumes, self-hosting requirements, and direct platform comparisons. When a user submits a broad query, CRUX generates additional predictive reply options covering the most common automation scenarios. Selecting any suggestion sends it as a message and preserves the full conversation history.

Contextual actions placed at the point of use

Share, log-in, and platform links are inserted into the conversation at the moment they become relevant. The share function becomes available only after CRUX has sufficient context and crosses a confidence threshold, so it never appears prematurely. Platform links go directly to each platform's page, putting the user's next action one tap away. Log-in is handled via magic link, no sign-up form, no password, no friction.

Every interface decision was evaluated against one criterion: does this reduce the number of steps between the user's goal and a correct recommendation.

Screenshot showing the landing page input field with pre-populated example queries and quick reply options generated mid-conversation — Predictive prompt suggestions and quick replies

Product Architecture

Conversation and report as two distinct output types

The conversation and the shareable report serve different functions. Treating them as two views of the same data would have compromised both.

The Conversation

Adaptive and session-bound

CRUX surfaces pricing assumptions, explains trade-offs, and adjusts based on follow-up. Each session updates as the user provides more context.

The Report

Structured and permanent

A public page with a readable URL, timestamped pricing data, and source links. Structured as Research Question, Key Findings, Recommendation, and Details. Indexed and shareable.

Recommendations are produced by a deterministic scoring model, not by the language model. The language model extracts intent and formats output. Ranking runs across four weighted dimensions: relevance, budget fit, technical complexity, and integration compatibility. This separation makes recommendations auditable and keeps affiliate relationships structurally outside the scoring logic.

The shareable public report page showing timestamped pricing data and source links — The shareable public report with timestamped pricing data and structured findings

Outcome

Scope of work delivered

CRUX launched as a live product at crux.do. Full scope delivered: conversational AI advisor, structured comparison reports, complete brand identity system, and UI component library. The platform runs on a weekly data verification cycle across all tracked automation platforms. Pricing data is timestamped and source-linked on every report.

Key Observations

Three conclusions from building CRUX

Conversational AI performs better with structure, not freedom

A blank input is not a neutral design decision. Defining the question sequence and providing reply options produced measurably better engagement than an open-ended prompt. Structure reduces error and improves the quality of inputs the system has to work with.

Separating the AI layer from recommendation logic is a product requirement

Using the language model for intent parsing and output formatting, while running recommendations through a deterministic scoring model, made the product auditable and commercially credible. It also made bias structurally harder to introduce.

A conversational interface is not a substitute for UX thinking

Replacing a traditional UI with a chat window does not reduce the designer's responsibility to the user. On CRUX, every step required deliberate decisions about what information to show, when to show it, and what the user should never have to ask for. Cognitive load applies equally to conversational products. The interface has to anticipate user needs, not wait for them to be articulated.

CRUX: An AI advisor that does not make the user do the work