Human Factors Led User Experience
Proven in industries that could not afford to guess.
The stakes differ; the errors do not.
Aviation, rail, energy and medicine settled decades ago how to show that a design is safe to put in front of a person. The methods they built are not exotic, and none of them requires a reactor; they apply as well to a checkout flow, a dashboard or an onboarding sequence, because the person on the other side is the same person.
The position
Human factors trades easy answers for better questions.
Proven is not the same as certain. Human factors does not turn taste into arithmetic, and anything sold on that promise is something else.
What it offers is narrower and more useful: where a design is likely to fail, roughly how often, and what would have to change. It replaces confident assertion with stated uncertainty, which looks like less and is worth considerably more.
Writing
Arguments worth having
Working notes on making design evidence hold up, for anyone who has to defend a decision to someone who does not care how it looks.
Training is a control of last resort
Every industry required to prove its designs safe ranks remedies in the same order, with instruction near the bottom. Tooltips, walkthroughs and confirmation dialogues all belong to that tier.
A journey map records intent, not action
A journey map records what a team intends a person to do; a task analysis records what the person does, including the checks, waits and recoveries the map leaves out.
Automation is a change of task, not a removal of one
The easy half of a task is taken and the residue left behind, to somebody who is now out of practice at the moment they are most needed. Bainbridge described the shape of it in 1983.
From the literature
Four findings the field settled long ago
None of this is new: it is decades old, well evidenced, and largely absent from the way digital products are designed.
Human error is a consequence, not a cause. The useful question is never who was careless, but which conditions made that error the likely thing to happen.
Automating the easy parts of a task leaves the person holding the hard residue, and out of practice at the moment they are most needed.
Under cost and effort pressure, work drifts steadily towards the boundary of acceptable performance, and nothing fails until an ordinary variation reaches a system already at the edge.
The five-user figure comes from a model that assumes problems are independent and equally easy to detect; rare, unevenly detectable failures break both assumptions.
Paraphrased rather than quoted, and each checked against the source.
Fitts publishes the first allocation-of-function list, setting what machines do better against what people do better. Every automation decision since has been a version of that question.
Annett and Duncan publish hierarchical task analysis. It remains the starting point for most of the methods in this library.
Bainbridge describes the ironies of automation: the easy half of a task is automated and the residue left to the person. The paper reads as a description of AI copilots.
Methods in the library, each with the question it answers, the work it came out of, and what it leaves unsettled.
Resources
Ready to use today
The Error Modes Checklist
Fourteen ways people predictably fail at an interface, each with the design question that catches it. It runs to one page, built to sit open in a design review.
The HF → UX glossary
Fifteen terms that do real work, translated out of the literature into language a product team will use. No email address is required.
Eight methods, two ways in
Indexed by method and by the problem a team is facing, each entry attributed to its origin and scoped to what it does and does not settle.
Occasional notes, when there is something worth saying
New writing, new method pages and templates as they are finished, with no fixed schedule, no funnel and no newsletter about the newsletter.
Library
The methods, and what
each one proves.
Each of these has been standard practice in safety-critical engineering for decades, proven against consequences most product teams never have to face. None of it is new, and none of it is out of reach. Each entry sets out the question the method answers, where it came from, and what drives the cost of running it.
The library is indexed two ways: by method, and by the problem a team is facing.
The methods
Hierarchical task analysis breaks work into what a person does, in order, to the level of detail at which failure becomes visible. It records the checks, waits and recoveries that a happy-path journey omits, and most of the other methods here start from it.
- Answers
- What are people doing, and where is it fragile?
- Feeds into
- Error prediction, test protocol design, requirements.
- Origin
- Annett & Duncan, 1967. Standard in process industries since.
- Scales with
- The number of tasks that are critical, and whether they can be observed rather than described.
Related: A journey map records intent, not action.
Each step is examined systematically for the forms of error it invites: omission, wrong selection, the right action on the wrong object, action too early, action repeated. Each is rated, and each is mitigated in the design rather than in the training material.
| Step | Error mode | Consequence | Risk | Design mitigation |
|---|---|---|---|---|
| 2.2 | Adjacent unit selected | Plant started with crew on site | High | Live permit state blocks selection; units shown by location, not ID alone |
| 2.3 | Valve state misread | Flow into an isolated section | High | Show measured state, not commanded state; disagreement is made loud |
| 1.0 | Step omitted under interruption | Permit left open | Medium | Resumption cue on return; incomplete state visible from the overview |
| 3.0 | Premature closure of check | Error passes verification | Medium | Independent re-entry rather than confirmation of what is shown |
- Answers
- What will go wrong, how badly, and what stops it?
- Techniques
- SHERPA, HEART, TRACEr, FMEA adapted for human action.
- Origin
- Nuclear and process safety, 1980s.
- Scales with
- The depth of the task analysis underneath it; without one, most of the work remains to be done.
Related: An error message is a system report, not a recovery instruction.
Allocation of function decides, deliberately, which work the machine does and which stays with the person, and then checks that what remains is a coherent job rather than the residue nobody wanted. The question sits underneath every automation and every AI feature, and is more often answered by accident than by design.
- Answers
- What is the person being asked to do, and can it be done well?
- Watch for
- The ironies of automation: the easy half taken, and the hard residue left to a person who is now out of practice.
- Origin
- Fitts, 1951. Sharpened considerably by Bainbridge, 1983.
- Scales with
- How much of the decision has already been taken elsewhere; the analysis is rarely what takes the time.
Related: Automation is a change of task, not a removal of one.
Attention is a budget that every interface spends, most of them without noticing. Workload measurement shows whether a design saved effort or moved it somewhere less visible; situation awareness measurement shows whether the person knows what the system is doing, which is the question behind many automation failures.
- Answers
- How much does this cost the user, and do they know where they are?
- Instruments
- NASA-TLX, SART, SAGAT freeze-probe, secondary-task methods.
- Origin
- Aviation human factors, 1980s–90s.
- Scales with
- Whether it can be attached to a study already planned; run on its own it costs considerably more.
A mental model is what the user believes the system is doing, as distinct from what it is doing. Errors cluster where the two diverge, so a large share of them can be predicted by eliciting the model rather than waiting for the mistake.
- Answers
- Where will people be confidently wrong about the system?
- Techniques
- Teach-back, prediction tasks, card sorting against system state, critical decision method.
- Origin
- Craik and Johnson-Laird, by way of Norman.
- Scales with
- Access to people who use the system in earnest, rather than people who resemble them.
Related: A persona is an audience description, not a user model.
The cognitive walkthrough is a structured expert evaluation that takes a task step by step and asks at each one whether the user would know what to do, would see the control that does it, and would know that it had worked. It is disciplined enough to be repeatable and cheap enough to run before anything is built.
- Answers
- Can a first-time user get through this without being told?
- Best for
- Onboarding, infrequent tasks, and any task with no chance to learn by repetition.
- Origin
- Polson, Lewis, Rieman & Wharton, 1992.
- Scales with
- The number of tasks and evaluators; a second evaluator costs less than is usually assumed and finds more.
Related: Usability testing samples what is noticed, not what occurs.
Evidence design settles, before the study, what result would change the decision and whether the planned study could produce it. A good deal of UX research fails at this point rather than in fieldwork: the sample cannot support the claim, or the claim was never stated precisely enough to be wrong.
- Answers
- What can this study prove, and is that worth the cost?
- Covers
- Sample justification, effect size, task representativeness, acceptance criteria set in advance.
- Origin
- Experimental psychology, borrowed directly.
- Scales with
- Little: it is the cheapest method on this page and the most reliably skipped.
Related: Intent and behaviour are two research questions, not one.
The evidence record is the written trail from intended use, through the failures predicted and the design decisions taken, to what testing showed. Regulated teams are obliged to keep one; others discover the need the first time something goes wrong and nobody can reconstruct why the design is as it is.
- Answers
- Can the design be defended in a year, to someone hostile?
- Covers
- Intended use, predicted failures, decision rationale, test results, open risks.
- Origin
- Safety cases in nuclear and process industries.
- Scales with
- When it is started. Kept as the work proceeds it is barely visible; reconstructed at the end it is brutal.
Problems a team is facing
Onboarding & account creation
People fail once, silently, and do not return; the analytics record a drop-off without a cause.
Forms & data entry
Wrong values are entered with confidence, and the dangerous ones are plausible, accepted and wrong.
Error messaging & recovery
The message explains what the system thinks happened, rather than what the person should now do about it.
Checkout & payments
Amounts, currencies and recipients are irreversible once sent, and notoriously sensitive to a dropped decimal.
Dashboards & monitoring
Everything is on screen and nobody can say what the system is about to do; data on display is not the same as awareness.
Workflow & internal tooling
Small frictions are multiplied by thousands of repetitions, in work nobody outside the team has watched being done.
Safety-critical interfaces
The mitigation in place is often a warning label, a training slide and the hope that nobody is tired.
AI-assisted interfaces
Automation takes the easy half and leaves the person to catch what it gets wrong, while out of practice at catching anything.
Accessibility
Conformance can pass while the task remains impossible: compliance measures the interface, and human factors measures whether the work gets done.
Nine problems and eight methods so far; work in progress is listed on the resources page.
Sequence
How the methods fit together
Scoping the risk
The question is which tasks, done by whom, carry a consequence of failure that is real. It is short work, and it can reorder a roadmap on its own.
Decomposition and observation
Task analysis is done against observed use rather than the flow diagram; where observation is not possible, that is stated and the confidence marked accordingly.
Failure prediction
Systematic error identification produces a ranked list of what will go wrong and the design changes that prevent it.
Targeted testing
The test protocol is aimed at the predicted failures, with acceptance criteria agreed before anyone runs a session.
The record
The work is written up to outlast the team that did it, so that a later reviewer, auditor or investigator can follow the reasoning.
Resources
Ready to use today.
Templates, checklists and definitions, free of charge and gated by nothing more than an email address where one is needed. Each is labelled honestly: what exists is here, and what does not yet exist says so.
Available now
Ready to download
The Error Modes Checklist
Fourteen ways people predictably fail at an interface, from mode error and capture error to premature closure, each with the design question that catches it. It runs to one page and is built to sit open in a design review.
The HF → UX glossary
Fifteen terms that do real work, translated out of the literature into language a product team will use. It sits below on this page, with no email address required and no download to lose.
Two rather than twenty: everything on this page should be fit to hand to a client without apology.
Glossary
Fifteen terms worth knowing
Human factors has a vocabulary, much of it more precise than its UX equivalent. These are the terms that change how a screen is read.
In progress
Being written
Listed as notice of what is coming rather than to fill the page; each is released when it is good enough to use unsupervised.
Notice when they are finished
A single email when something new is finished, with no fixed schedule and no funnel.
Integration
Where human factors fits into an engineering programme.
Human factors usually fails in a predictable way: the analysis is late rather than bad. By the time the report lands, the architecture is frozen, the headcount is already in the business case, and the only move left is to write a training requirement and call it a mitigation.
Scope
Five questions a programme has to answer about people
Different industries give these different names and different paperwork. Underneath, every serious programme answers the same five, and those that skip a question tend to meet it again later at considerable expense.
Who is the system for?
The real population who will operate and maintain it, rather than a persona: what they will know on day one, who can realistically be recruited and kept, and how that drifts across the life of the system.
What is the person being asked to do?
This is allocation of function: which work goes to the machine, which stays with the person, and whether what is left is a coherent job or residue nobody designed on purpose.
What has been delegated to training?
Every decision the design declines to make is handed quietly to training, to procedure, or to the operator’s judgement on the day. The question is which ones, and whether that is a defensible place to have put them.
What happens when someone gets it wrong?
Predicted error modes, their consequence and whether recovery is possible make up the human contribution to the failure argument, in whatever form the sector requires and for whoever ends up reading it.
What conditions does the design assume?
Staffing, handover, communication, time pressure and the organisational context on which the design silently depends. In-service problems are usually born here, years before anybody notices them.
Lifecycle
Where the work belongs
The earlier the work is done, the cheaper it is, and the margin is not a narrow one. Whatever a programme calls its gates, the shape is the same.
Concept
What the system is for, who it is for, and where people-related risk sits; the cheapest human factors work a programme will commission.
Requirements
That understanding is turned into requirements one party can build to and another can verify, and most of the value is captured or lost here.
Design
Allocation of function, task analysis against real use, interface specification and workload, all of them expensive to revisit later.
Verification
Evidence is judged against the criteria set at requirements, not against whatever the test happened to show.
In service
Incidents, near misses and workarounds are the richest data a programme will hold, and usually the least read.
Change
Every modification reopens the argument, which many programmes discover one change too late.
Where it goes wrong
Three ways it usually fails
Late analysis is expensive theatre.
By detailed design, most of the decisions this work exists to influence have already been taken. Two weeks at concept is worth more than two months at design review spent explaining what can no longer be changed.
A deliverable is not an outcome.
The work can be compliant in every particular and entirely inert: every document produced on time, and nothing about the design different. Its purpose is to change decisions, and an argument serves that purpose better than a document nobody opens.
Training is a control of last resort.
Where a design puts an error-prone step in front of a person and the remedy offered is more training, a stage has been skipped. Training is occasionally the right answer; more often it is the cheapest thing to write down.
Writing
Method, argued.
Working notes on making design evidence hold up, written for anyone who has to defend a decision to someone who does not care how it looks.
Download
The Error Modes Checklist
Fourteen predictable ways people fail at an interface, each with the design question that catches it. It runs to one page, built to sit open in a design review.
Get the checklistTraining is a control of last resort
Risk-control standards rank warnings, manuals and training last, and agree on why: such measures work only where the instruction was retained, is to hand at the moment of decision, and has not decayed in between.
A journey map records intent, not action
A journey map records what a team intends a person to do; a task analysis records what the person does, and can be interrogated where the map can only be presented.
Intent and behaviour are two research questions, not one
Discovery establishes what somebody means to do. Whether they then do it is a separate finding, and the research that produces it looks nothing like an interview.
Usability testing samples what is noticed, not what occurs
A test session produces a record of two acts of noticing: the participant’s and the observer’s. Neither is a complete record of what happened, and the gap between them has been measured.
A persona is an audience description, not a user model
Personas communicate who a product is for, and do it well. The thing they are routinely asked to do instead — stand in for a specification of human capability — is a job they were never built for.
An error message is a system report, not a recovery instruction
The standard that governs error handling asks for three things, in order: that the error be made impossible, that the system survive it, and that it be reversible. The message appears in none of them.
A confirmation prompt tests attention, not intention
Twenty-five million warnings were measured in the field. The best-performing ones stopped most people; the worst were bypassed seven times in ten. What none of them could do was tell a correct decision from a confident mistake.
Conformance is an interface property, not a task outcome
WCAG defines conformance for pages. Whether a person can finish the job the pages exist for is a different measurement, and the two have been shown to come apart.
Automation is a change of task, not a removal of one
Automating a step does not leave a person with one fewer thing to do. It leaves them with a different thing to do, usually monitoring, which is the task people perform worst.
Migration toward the boundary is a normal process, not a deviation
Workarounds are not indiscipline. Rasmussen described the pressures that move any working system toward the edge of what it can tolerate, and they operate whether or not anyone intends them to.
Human error is a consequence, not a cause
Closing an incident as user error terminates the investigation at the point where the useful part begins. The label describes the outcome, and offers nothing about the conditions that produced it.
Mental strain is a measurement, not an impression
A design critique calls an interface heavy. Ergonomics declines to treat that word as technical at all, and defines two quantities underneath it that can be measured and compared.
Twelve pieces, each with its sources listed, its paraphrasing labelled, and its standards cited to the clause. Corrections are welcome and are answered first.
About
A working library, built in the open.
Human Flux exists to move a set of methods across a border. Human factors has spent seventy years working out how to design for people who are tired, interrupted and under pressure, and digital product design has, for the most part, not read it.
This site is the translation: the methods written up so that a product team can run them, with the sources kept visible so that the working can be checked.
What is here
Three things, and no sales pitch
Eight methods, two ways in
Indexed by method and by problem, each with the question it answers, where it came from, and what drives the cost of running it.
Ready to use today
Checklists, templates and a glossary, free and honestly labelled: what exists is there, and what does not yet exist says so.
Arguments worth having
Twelve pieces on making design evidence hold up, aimed at anyone who has to defend a decision to someone unsympathetic.
The name
Why Human Flux
Human Factors, led. The initials are kept deliberately: HF is the discipline, and the discipline is the point.
Flux is the flow of people, attention and information through a system, which is what human factors studies: neither the interface nor the user, but the traffic between them.
Principles
What this site will be difficult about
Evidence has a grade
Every finding comes with what it rests on. “Two participants mentioned this” and “error rate of 11% across 34 trials” are both useful, and they will never be written down the same way here.
Design out rather than train out
A mitigation that consists of telling people to be careful is not a mitigation; it is an admission with a training budget attached.
The error is in the system
People fail predictably, in patterns documented for decades. A design that requires vigilance is a design that has decided to blame the user in advance.
Sources stay visible
Methods are attributed, findings are cited, and paraphrase is labelled as paraphrase. A site arguing for evidence has no business being vague about its own.
Authorship
Where this comes from
This is not a summary of the literature. It comes out of two decades of applying human factors in safety-critical industries — defence, energy and aviation — where the analysis had to hold up in front of a regulator, a safety case or an incident review.
What is written here is that practice translated for digital product work: the same methods and the same discipline, applied where the consequences are smaller and the errors identical.
Full credentials and worked case studies follow.
Contact
Questions, corrections, requests.
Nothing is for sale here, so this is not a sales channel. It exists for reporting that a method page is wrong, asking for one that does not exist yet, or putting a question to someone who has run these methods in practice.
info@human-flux.com
Replies usually follow within a few days. Corrections are answered first and credited on the page unless the sender prefers otherwise.