People decide in 50 milliseconds. Now your design process can too.
Everyone has an opinion about design. They use these products daily, so the opinion feels earned. But opinions aren't diagnosis. This is.
A 5-layer diagnostic framework grounded in cognitive psychology. Tells you which perceptual layer is failing, in what order to fix it, and why the fix works. Before anyone touches a pixel.
~100 peer-reviewed citations. 15 years applied. Open framework.
Created by Stefan Kovalik / Aurochs
The stakeholder says make the logo bigger. The VP doesn't like blue. The founder's wife thinks the hero needs a video.
The designer is either pushing pixels to someone else's vision or burning half their time persuading stakeholders why a bad instinct is bad. The person on the other end of the screen, how they actually perceive and decide in 50 milliseconds, is nowhere in the conversation.
PFD changes the conversation from opinions to diagnosis.
You look at a page and you see the layers. Not the colors, not the layout, not the copy. The perceptual gates between a human and a decision.
The hero is fine but L0 is broken: too many competing elements burning working memory before visitors reach the headline. L2 is leaking trust through inconsistent spacing the viewer's nervous system notices before the designer does.
You stop guessing. You fix L0 first because Lavie (1995) proved that lower-level processing demands eliminate higher-level processing capacity. Then L1. Then L2.
Each fix compounds upward through the stack, toward the thing every decision actually runs on: pre-verbal arousal. The viewer's nervous system fires before their analytical mind engages. They feel the outcome before they evaluate the offer.
Damasio (1994) showed the body generating emotional signals that bias decisions before conscious deliberation. LeDoux (1996) mapped the pathway: 12ms emotional response, 200-300ms conscious evaluation. The feeling arrives first. Always.
Design becomes a lens of empathy to solve with, not a set of business objects to align with the user. The stakeholder can still disagree, but now they're disagreeing with cognitive psychology instead of someone's taste.
LLMs miss things. Not sometimes. Every time. A single pass through any problem is probabilistically incomplete, the same way a 3D renderer's first pass produces a noisy image. The fix is the same too: run it again. Each pass catches what the last one missed.
The Ralph Loop applies this to PFD. Run a design, plan, pitch, or system through the 5 layers. Loop until exit conditions fire and completion signals hold up under evaluation. Brute force repetition against your own instructions, compounding toward an asymptote.
Round 1 catches the obvious layer violations. Round 2 catches what only surfaces after the obvious ones are fixed. By round 3, the conversation isn't about opinions. It's about which perceptual layer the remaining issues live on and what the science says to do about them.
Match depth to stakes. Quick design decision: 1 pass. Feature spec: 5. Landing page or pitch deck: 10. Systems architecture or deep research: 50. I've run 100+ on things that needed it. Each pass costs more and returns less, so you stop when confidence stops climbing.
Fix bottom-up. Upstream failures block everything downstream.
| Layer | Gate | What Breaks If It Fails |
|---|---|---|
| Cognitive Load (L0) | Working memory: 3-5 chunks | Visitor leaves before processing anything. Bandwidth consumed by noise. |
| First Impression (L1) | 50ms visual verdict | Attention never activates. Everything downstream multiplies by zero. |
| Processing Fluency (L2) | Easy to process = feels true | Trust erodes subconsciously. The visitor can't explain why they don't believe you. |
| Perception Bias (L3) | Users autopilot, rationalize after | You're designing for what users SAY instead of what they DO. The gap is where conversions die. |
| Decision Architecture (L4) | Structure shapes choice | The trail is broken. Interested visitors can't find the path to action. |
Full framework (700+ lines, ~100 citations): framework/PERCEPTION-FIRST-DESIGN.md
Clone the repo. Read the framework. The five layers are a diagnostic instrument: walk any design, email, pitch, or onboarding flow through them bottom-up. You'll see what's broken and why, grounded in the same prediction error logic the brain runs on.
git clone https://github.com/skovalik/perception-first-design.gitAutomated design thinking. The PFD skill runs on anything humans perceive: interfaces, landing pages, emails, ads, pitches, onboarding flows, documentation.
Install via your own marketplace (no review required):
/plugin marketplace add skovalik/perception-first-design
/plugin install perception-first-design@perception-first-designAfter install, use the skill in four ways:
# Bare invocation — auto-detects mode from input
"PFD this landing page: https://example.com" # → evaluate (URL = artifact)
"Run PFD on our onboarding flow design problem" # → solve (problem = derivation)
"What happens if we remove the URL bar from Chrome?" # → analyze (hypothetical)
# Force Mode 2 (solve): derives a solution from cognitive constraints
/perception-first-design:solve "Should we use a modal or inline expansion for the pricing FAQ?"
/perception-first-design:solve "We're losing mobile users at checkout. What's failing?"
# Force Mode 1 (evaluate): corpus-backed audit of an existing artifact
/perception-first-design:evaluate https://example.com
# Force Mode 3 (analyze): descriptive cascade for hypotheticals and mechanism questions
/perception-first-design:analyze "What happens if Twitter removes the like button?"
/perception-first-design:analyze "Why does dark mode feel premium even when contrast is worse?"
# Composite: run all three modes on the same input
/perception-first-design:all "We're considering removing the URL bar. Run the full PFD treatment."Mode 1: Evaluation. Walk an existing design through the 5 layers. Each layer validates or flags. Works on URLs, screenshots, mockups, HTML files, copy, emails, pitches. Anything humans perceive.
Mode 2: Derivation. Work bottom-up through the 5 layers. Each layer produces a hard requirement. The solution emerges from accumulated constraints, not intuition, not competitive copying, not what the HiPPO wants. This is the mode that generates answers you wouldn't have reached by instinct.
Mode 3: Analysis. Walk the 5 layers as predictive lenses, not as constraints. Output cascading consequences, trade-offs, and integrative compounds. Good for hypothetical change questions ("what happens if X"), mechanism questions ("why does X work"), and behavioral observations ("users say X but do Y"). Produces results, not recommendations. The cognitive contract is descriptive, not prescriptive.
Rule Zero: do not propose any solution until all 5 layers are analyzed. The right answer is the only one that survives all five filters. (Analysis follows a parallel rule: walk all 5 layers as descriptive lenses, then check trade-offs and integrative compounds, before treating the cascade as complete.)
Here is the full /pfd:all output on a thought experiment about
browser design. The question hits all five perceptual layers in
different ways and surfaces a finding the v0.6.0 solve mode
compressed to a single verdict.
/pfd:all on "What happens if we remove the URL bar and browser tabs from Chrome?"
Three lenses, by cognitive contract.
5 layer-cascade consequences, 3 integrative compounds.
Users have to hold "where am I" and "what tabs were open" in their heads instead of in the UI.
The URL bar and tabs are memory prostheses (location + parallel-task tracking). Removing both moves both into the head simultaneously. Power users running 5+ context-switching tasks hit the 4-5 chunk Cowan ceiling within a session.
What happens: Light users do not notice for weeks. Power users hit the ceiling immediately, drop context, churn first. Tab groups (the hierarchical memory structure power users built workflows around) compound the loss.
Citations: Cowan (2001, 2010) for working memory capacity ~3-5 chunks; Hassin et al. (2009) for WM cost of unconscious processing.
The 50ms aesthetic credit reverses asymmetrically: casual users keep it, navigation-heavy users lose it instantly, touchscreen users hit a hard wall.
Three populations split immediately. Casual content consumers (social, news, single-page apps) keep first-impression high. Power users lose it within 30 seconds. Touchscreen-only users hit catastrophic failure (hover-discoverability does not exist on touch). Plus a novice cascade: cannot share URLs verbally because there is nothing to copy.
What happens: Net first-impression bimodal. Touchscreen users churn first, novice users second, casual content consumers stay loyal.
Citations: Lindgaard et al. (2006) for 50ms visual appeal; Kurosu & Kashimura (1995) / Tractinsky (2006) for aesthetic-usability effect.
3. Predictive processing failure cascades into the extension and automation ecosystem [Processing Fluency]
25-year-deep convention breaks AND every extension, password manager, and automation script that assumes URL/tab DOM exists breaks too.
Predictive processing breaks (the schema users pre-load in 200ms is gone). Safari's precedent shows users CAN adapt when only one signal changes; this proposal changes both. Adjacent infrastructure: extensions, automation, password managers all break. Password managers throwing errors because they cannot read the URL is the most dangerous version.
What happens: Sessions feel 15-25% slower during transition. Trust erodes laterally. Extension breakage generates a separate "something broke" cascade misattributed to Chrome instability rather than the design change.
Citations: Reber & Schwarz (1999) for fluency-truth effect; Clark (2013) for predictive processing; Forster, Leder & Ansorge (2013) for felt vs measurable fluency.
Surveys say "I love clean interfaces." Analytics show users compensating with more windows, heavier bookmarks, search-as-URL-bar. PM dashboard reads "NPS up, engagement flat" for 6 months while power users churn invisibly.
NPS goes UP (clean things survey well). Session count goes UP (workflow fragmentation reads as engagement). Pages-per-session goes DOWN (misattributed to engagement quality). Task completion time goes UP (rarely tracked at right granularity).
What happens: "NPS up, engagement flat" looks fine. PMs misread for 6 months until cohort retention shows power-user churn. By then rollback is politically expensive.
Citations: Nisbett & Wilson (1977) for System 2 rationalization; Kahneman (2011) for autopilot decisions.
5. Substitution reshapes the trust surface and surrenders identity verification [Decision Architecture × Perception Bias]
Without URL bar and tabs, navigation routes through search engines. Chrome donates query share AND loses the visual surface that anchored phishing detection. Phishing volume drops, severity rises.
Self-tax: Chrome donates 1-3% query share to search partner. Phishing trade-off: volume drops (lookalike-URL spoofing loses primary surface) but per-attempt severity rises (no surface to verify chase.com.evil.io). Regulatory cascade: EU DMA/DSA may classify URL display as regulated consumer-protection surface.
What happens: Search partner gains query share. Identity-sensitive brands escalate to Google. EU/UK/AU/CA regulators file inquiries within 6 months. Banks run "always verify the URL" campaigns and discover the URL is no longer visible.
Citations: Trope & Liberman (2010) for construal-level theory; Seckler et al. (2015) for trust as conjoint perceptual judgment.
Power-user tech press in days. Enterprise security in weeks. Regulators in months. Each wave narrows Google's response window.
Wave 1 (days): Hacker News, Reddit, tech press. Migration to Firefox/Brave/Arc. Wave 2 (weeks): enterprise IT, security teams. Wave 3 (months): regulators, academic citation infrastructure.
What happens: Backlash compounds louder than any single layer. Google's response window narrows; by Wave 3, regulatory action constrains future design freedom regardless of preference.
Citations: Applied prediction; no direct framework citation. Closest adjacent: Seckler et al. (2015) for trust as a conjoint judgment that aggregates across users.
Once users build search-as-navigation habits over 4-6 weeks, rollback recovers WM and fluency but cannot recover the donated query share. Asymmetry is permanent.
Rollback recovers WM and fluency cost. Does not recover habits. Some users never re-adopt URL-bar use because their workflow stabilized around bookmarks/history.
What happens: Even successful rollback leaves ~0.5-1% query share permanently donated. The asymmetry is why the change is high-risk regardless of outcome.
Citations: Hertwig & Erev (2009) for description-experience gap.
Chrome OS, Android Chrome, embedded WebViews, Edge/Brave/Arc/Vivaldi all face follow-or-fork. Edge gains positioning advantage by NOT removing the URL bar.
Embedded WebViews carry the change forward inside other apps. Chromium derivatives must follow (inherit backlash) or fork (absorb maintenance). Chrome OS becomes harder to use as navigation OS because the URL is the shell.
What happens: Edge gains relevance positioning as "the one that didn't remove the URL bar." Some Chromium derivatives fork the URL-bar code permanently, fragmenting upstream-merge.
Citations: Applied prediction; no direct framework citation. Strategic/ecosystem cascade rather than perceptual finding.
R1 [Cognitive Load]: The browser MUST externalize location and parallel-task tracking somewhere visible. WM cannot hold either. Citations: Cowan (2001, 2010); Hassin et al. (2009).
R2 [First Impression]: Substitute affordances MUST be discoverable in the first 50ms across all device types. No hidden gestures or hovers without visual telegraph. Citations: Lindgaard et al. (2006); Tractinsky (2006).
R3 [Processing Fluency]: If we break the address-bar-on-top schema, replace it with one equally consistent and predictable. URL/tab DOM/API surfaces MUST remain accessible programmatically. Citations: Reber & Schwarz (1999); Clark (2013).
R4 [Perception Bias]: Measure behavior, not survey claims. Test by behavioral metrics, not stated preference. Citations: Nisbett & Wilson (1977); Kahneman (2011).
R5 [Decision Architecture]: Replacement MUST preserve identity verification (anti-phishing surface), navigation control, and parallel-task management. Trust surface stays visible. Citations: Trope & Liberman (2010); Seckler et al. (2015).
Primary solution: Hide-on-idle, surface-on-approach. URL bar and tabs auto-hide while reading content. Surface on cursor approach (desktop) or edge swipe (touch). All five R's satisfied.
Alternative 1: Vertical rail. Persistent left sidebar, schema shifts horizontal to vertical, functions remain.
Alternative 2: Move tabs to the bottom; keep URL bar. Lower-risk, smaller aesthetic gain.
Original proposal hard-fails R1, R3, R5.
Gap: Touch users underserved by all alternatives. The non-obvious solution: hide-on-idle plus a thin always-visible edge strip that expands on tap. Satisfies all five R's including the touch population.
Skipped. Evaluate requires an artifact (URL, screenshot, HTML, copy of the proposed minimal browser). Attach a mockup of any alternative to score against the layer rubric.
Initial findings across three lenses. Ralph Loop any consequence/requirement, cite further, dive into a single mode, or ask any follow-up to dig deeper.
What each section delivers:
Analyze surfaces 5 layer-cascade consequences plus 3 integrative compounds. Each consequence has a plain-language italic subtitle, mechanism prose, a "What happens" prediction, and 1-3 strongest supporting citations from the framework's research base.
Solve derives R1 through R5 from the layer constraints, then proposes solutions that satisfy all five (hide-on-idle, vertical rail, move tabs to bottom). The literal "remove both" proposal hard-fails three requirements. That hard-fail is the answer to whether the original idea is good.
Evaluate skips because no artifact was provided. Attach a mockup of any alternative to score it against the layer rubric.
The output ends with one italic line inviting deeper paths: Ralph Loop a finding, cite further, dive into one mode, or ask any follow-up.
.claude-plugin/ Plugin manifest + marketplace catalog
plugin.json Metadata (name, version, license, keywords, path overrides)
marketplace.json Self-hosted marketplace entry pointing at this repo
framework/ The complete PFD framework (v3.6, ~100 citations)
PERCEPTION-FIRST-DESIGN.md
ADHD-CURB-CUT.md ADHD and autism as diagnostic instruments, not disclaimers
corpus/ The evaluation engine
core/ Rubric, constraints, psychology reference, output schema
heuristics/ 26 rules across 5 layers with psychology citations
design-systems/ Tailwind, WordPress, Shopify detection profiles
worked-examples/ 7 calibrated examples from terrible (18/100) to excellent (92/100)
skills/pfd/ The PFD skill (derivation protocol, layer summaries)
references/learnings/ 29 curated learnings sharded by primary layer
commands/ /perception-first-design: solve, evaluate, analyze, all
scripts/ gen-pfd-index.py (regenerate _index + _search from atoms)
The framework is open. The calibration is earned.
This repo ships 29 curated learnings as atom files under
skills/pfd/references/learnings/, organized by perceptual layer (L0/,L1/,L2/,L3/,L4/,meta/,cross/) with a generated_index.mdand_search.jsonfor lazy-load.Contribute a learning via pull request or the Learning Submission issue template.
I designed this framework with autism and ADHD as diagnostic instruments. The cognitive constraints I design around are constraints every user has. Mine are just louder.
A page that overloads an ADHD user's working memory is overloading everyone's. The ADHD user just notices first. Same principle as curb cuts: built for the constrained case, better for everyone.
Full write-up: framework/ADHD-CURB-CUT.md
Make Me Think: Perception-First Design for the Post-Usability Era by Stefan Kovalik. 12 chapters at aurochs.agency/writing/make-me-think/. The framework is the reference; the book is the story.
Kovalik, S. (2025-2026). Perception-First Design: A cognitive
psychology framework for design. Version 3.6. CC BY-SA 4.0.
https://github.com/skovalik/perception-first-design
Machine-readable: CITATION.cff
CC BY-SA 4.0. See LICENSE.
Practice exemption: applying PFD in practice — running analyses, producing recommendations, delivering client work, designing products with PFD principles — does not create a derivative work and is not subject to share-alike. You can apply PFD commercially without any obligations on your work product. Only redistributions or modifications of the framework text itself trigger share-alike.
"Perception-First Design" is a trademark (USPTO Serial 99686343). The methodology is open and I want it used and shared. Please credit it.
If you want to use the name commercially (a product, service, certification, or branded offering), just email me at stefan@aurochs.agency. I'm very likely to say yes. I mostly want to know who's using it and how.
Attribution: Perception-First Design™ by Stefan Kovalik / Aurochs.
See NOTICE for full terms.
See CONTRIBUTING.md. Key rule: every claim needs a citation or must be labeled "practitioner observation."
Stefan Kovalik. BA Psychology, UC Santa Barbara. 15 years design and development. Autism and ADHD as diagnostic instruments, not disclaimers.