Skip to content

Research-backed design ​

Guildhall borrows from research and practice, but it does not pretend a paper can make product decisions for you. The research helps explain why the product is shaped the way it is: smaller tasks, bounded questions, visible uncertainty, review loops, and an audit trail you can actually read.

This page is the backing shelf. It is not a manifesto, and it is not a claim that Guildhall has solved the hard parts of AI-assisted software work. It is a plain map from useful ideas to product choices.

Cognitive load ​

Source: John Sweller's work on cognitive load theory, including Cognitive Load During Problem Solving: Effects on Learning, and the wider human-factors tradition around limited working memory, including George Miller's The Magical Number Seven, Plus or Minus Two.

Accessible explainer: The Interaction Design Foundation's What is Cognitive Load? connects cognitive load to interface design: high mental effort makes products harder to understand, remember, and use.

What Guildhall takes from it: a giant wall of context is a bad way to ask for a decision. The app tries to show the smallest useful question, the reason it matters, and the evidence behind it.

Where it shows up:

  • one-question-at-a-time Pressure-Test Intake;
  • compact task cards with deeper drawers nearby;
  • Journey before raw Transcript on completed tasks;
  • summary-plus-evidence references instead of dumping every receipt into the first screen;
  • task sizing, so a worker and reviewer can hold the unit of work in their heads.

Goal setting ​

Source: Edwin Locke and Gary Latham's Building a Practically Useful Theory of Goal Setting and Task Motivation.

Accessible explainer: MindTools has a plain-language overview of Locke's Goal-Setting Theory that frames useful goals as clear, challenging, committed-to, feedback-rich, and realistic enough to act on.

What Guildhall takes from it: clear goals change behavior. Vague "do your best" prompts make it easier for an agent to sound busy while drifting away from what you needed.

Where it shows up:

  • task blueprints with goals, non-goals, acceptance criteria, and checks;
  • request-shape detection before a New Request becomes work;
  • separate learning/exploration from implementation when the goal is still uncertain;
  • review against the accepted plan, not against how confident the transcript sounded.

Flow and work in progress ​

Source: The Kanban method's emphasis on visualizing work, limiting work in progress, and managing flow; see the Official Guide to the Kanban Method.

Accessible explainer: The official guide explains Kanban in practical terms: make work visible, limit work in progress, manage flow, and improve the system over time.

What Guildhall takes from it: more active work is not automatically more progress. Too many open lanes make blockers easier to ignore and handoffs harder to trust.

Where it shows up:

  • finishing active implementation, review, and gate work before grabbing more;
  • visible statuses instead of a vague "running" blob;
  • task sizing and split recommendations;
  • blocked work that stays visible instead of silently sinking under newer requests;
  • review budgets, so inspection depth can scale without inviting every possible reviewer every time.

Human-AI interaction ​

Source: Saleema Amershi and colleagues' Guidelines for Human-AI Interaction, validated through CHI 2019 design-practitioner studies and collected by Microsoft Research.

Accessible explainer: Microsoft's Eighteen best practices for human-centered AI design turns the research paper into product guidance about timing, uncertainty, correction, explanation, and recovery.

What Guildhall takes from it: AI systems need to make their abilities, uncertainty, mistakes, and recovery paths visible. When a system guesses, the user needs a way to inspect, correct, and continue.

Where it shows up:

  • clear routing for New Request before work starts;
  • explicit uncertainty and clarifying questions;
  • reframe before implementation, change orders after implementation starts;
  • reviewer findings, gate output, changed files, and context evidence attached to tasks;
  • Settings -> Memory suggestions that stay inspectable and reversible.

UX review calibration ​

Sources: Nielsen Norman Group's Ten Usability Heuristics and Usability Testing 101, Baymard Institute's Checkout UX 2025: 10 Pitfalls and Best Practices, the Deceptive Patterns taxonomy, and the FTC's Bringing Dark Patterns to Light summary.

Plain-language frame: The heuristics give Guildhall a broad checklist for legibility, control, consistency, recovery, and memory burden. Baymard gives it concrete examples of flow friction in high-stakes forms and checkout-like workflows. Deceptive Patterns and the FTC report give it names for the sharper edge cases: designs that hide, steer, delay, obstruct, or trap.

What Guildhall takes from it: UX review works better when it is calibrated against real failure cases. Guildhall checks for missing system status, internal jargon, weak recovery paths, hidden safe options, late-disclosed cost or risk, visual interference, and flows that make the reader's preferred path harder than the product's preferred path.

Where it shows up:

  • review calibration cases with hidden expected findings;
  • UX, accessibility, copy, trust, and deceptive-design review lanes;
  • screenshot, DOM, flow, and task-goal evidence requirements;
  • negative controls, so reviewers are not rewarded for finding imaginary problems;
  • review-plan budgeting, so high-risk flows get deeper review without making every task run every reviewer.

What this does not mean ​

Research does not decide your product taste. It does not know your release risk, your users, your codebase, or when a weird exception is worth it.

Guildhall uses these ideas as guardrails:

  • keep decisions small enough to answer;
  • keep work small enough to review;
  • keep uncertainty visible;
  • keep evidence attached;
  • keep recovery paths honest.

When those guardrails make the app feel too cautious or too loose, that is a product tuning problem, not a moral law from the literature.

Released under the FLL-1.2 License.