Jordan, Head of Security
45-person health-tech company, first SOC 2 audit booked for next month
"I don't need another checklist. I need to know the three things that will fail the audit, and fix them before Friday."
Persona voice, not a real quote
Every screen is a working prototype. Approve fixes, upload evidence, grant access that expires. All data is fictional.
A top navigation and a list-with-detail layout instead of the usual sidebar dashboard. Four connected pages and a command palette.
Open full screen ↗Clearpass is a fictional company that sells SOC 2 readiness to startups with 10 to 50 people. SOC 2 is a security audit that US business customers often ask for before they sign a contract. The buyer is usually one security lead, or an engineer who was given the job.
Business problem (assumed for this brief): accounts that stall below 50% readiness never book their audit, and churn before renewal. Design goal: get every new account past 50% in its first 30 days, without asking for risky access to production.
The client asked for three things: show exactly what is left, help people fix it fast, and make the evidence collect itself.
Security teams will not connect a tool that can change GitHub or AWS without asking. Anything that writes has to be asked for, scoped and logged.
Every action needs a person, a time and the exact change, in a form an auditor accepts.
Most buyers have no compliance team. Alerts and approvals have to work from a phone, between other jobs.
01 · What I owned
About 7 weeks, split into five phases. This is how I would plan the work as a real engagement, not a record of time spent.
02 · Who we designed for
Personas are working assumptions from desk research, not interview findings. They are what I would validate first.
45-person health-tech company, first SOC 2 audit booked for next month
"I don't need another checklist. I need to know the three things that will fail the audit, and fix them before Friday."
Persona voice, not a real quote
Owns the AWS and GitHub controls, receives the work
"Send me the exact change, not a compliance lecture. And don't touch production without asking."
Persona voice, not a real quote
03 · Research
Done: a teardown of Vanta and Drata public onboarding and control views, November 2026. This is the only research behind this concept.
If they won't, the fix flow falls back to "send the exact change to the owner as a ticket". The design already supports that, so the bet is cheap to lose.
04 · The core flow
Read-only links to cloud, code, identity and HR tools. Nothing can write by default.
Hourly tests turn every control into one readiness score. The prototype shows a trimmed set of 18.
Each failing control gets a prepared fix. You see exactly what will change, approve one scoped write, then it re-tests.
Evidence collects itself, shows how fresh it is, and reminds the owner before it expires.
The auditor gets a read-only, time-boxed seat with controls, evidence and the activity trail.
05 · Key decisions
06 · Roles and permissions
In the prototype you are Jordan, an Admin, so you can approve fixes. Owners prepare them. The auditor seat is designed but not prototyped.
Can: Connects systems, approves fixes that write, approves access (one of two for admin), manages members.
Can't: Cannot approve their own admin access request.
Can: Sees and prepares fixes for their controls, uploads evidence, snoozes non-critical alerts.
Can't: Cannot approve a write. A prepared fix goes to an Admin.
Can: Uploads the file they were asked for and comments on it.
Can't: No settings, no fixes, no access approvals.
Can: Read-only, time-boxed seat that ends with the audit window. Sees controls, evidence and the activity trail; can comment and request evidence.
Can't: Cannot see vendor contracts or change anything.
Scale beyond 200 seats (SSO, SCIM provisioning, bulk owner assignment, multiple workspaces) is in the v2 list below.
07 · Audit trail
Every row has a person, a time and the exact change. Rows are append-only, so nobody can edit history, and the auditor can export them as CSV.
Try it: in the web app above, open AC-03, approve the fix, and watch its Activity list update. The frame on the right is a static design of the auditor's view.
08 · Edge and failure states
Tests show "Not checked, GitHub disconnected 3h ago". Affected controls count as unknown, never as passing.
Steps stop at the failed one. Nothing is marked passing, the partial change is logged, and you get Retry and "Open in GitHub".
The prepared fix becomes a ticket for the control owner, with the exact change list attached.
Freshness shows days left. Owners get a reminder 30 days before expiry; expired evidence flips the control to "Needs evidence", in words.
Admin requests expire after 24 hours unanswered. The requester is told; nothing is granted.
No search results, everything passing, all caught up: each has one plain sentence instead of a blank panel.
09 · Accessibility
A target, not an audit: no external accessibility review has been done on this concept.
Every badge pairs a colour with a word: Passing, Failing, Needs evidence. Readable in greyscale and by screen readers.
Body text #475467 on white is 7.7:1, secondary text #667085 is 5:1. Filled buttons moved from teal #0E9384 (3.8:1 with white) to #107569 (5.6:1).
Every row is a real button. Ctrl K opens the command palette, Esc closes it, and inputs show a visible focus ring.
Mobile targets are at least 44px. Animations stop when the system asks for reduced motion.
10 · Design system
Navy #0C1A2B
Teal #0E9384
Mint #2ED3B7
Canvas #F6F7F9
Passing #079455
Failing #D92D20
Needs evidence #DC6803
Inter for everything people read. JetBrains Mono for control IDs, so they scan like the codes they are. Status badges always pair a colour with a word.
11 · Success metrics
This is a concept, so nothing has shipped and nothing has been measured. Baselines are assumptions to replace with real data in week one.
The north star, from the business problem
Matches the website promise
Counter-metric: speed must not cost trust
The renewal signal: people still use it once the audit is done
12 · What I left out of v1
13 · Reflection
I started with one-click auto-fix as the hero. Writing the permission model showed it was the riskiest idea in the product, so the fix became "prepared change, a person approves". Next time I would design permissions first, not the happy path.
Whether engineers trust a fix that comes from a compliance tool. If they don't, the better default is "send to owner as a ticket", and approval moves to the engineer.
5 security leads, unmoderated test of the Fix flow. Success: 4 of 5 approve the right change without opening docs.