AES Just Posted the New Technical Screen on HN: Show Us the PR, Not the Algorithm
A June 2026 HN "Who is hiring?" post from AES is the cleanest public template yet for screening AI-fluent engineers. Here is how to copy it.
If you ran a whiteboard round last week, you screened for a skill the candidate's tools have already replaced. The June 2026 Hacker News "Who is hiring?" thread made that explicit in a way no McKinsey deck has: a small environmental-test-chamber company posted a JD that publishes its exact stack, names Claude Code as its codegen layer, and tells applicants to send a PR instead of solving a puzzle. It is the cleanest public artifact yet of a screen that's been quietly mutating across the industry for 18 months.
The post that named the new bar
The anchor is a live listing from AES (associatedenvironmentalsystems.com) for a Summer Software Engineering Intern, remote in the US with occasional Boston days, $30 to $40 per hour, June 15 through August 22, 2026. The interesting part is not the comp band. It is the screen.
AES writes, in plain English in the JD itself:
The senior lead doesn't write code from scratch. Neither will you.
Then the bar:
The bar is NOT raw coding ability. We use Claude Code for all code generation. The bar is the ability to direct an AI at production engineering and catch it when it's wrong: schema literacy, diff-reading, end-to-end ownership.
The screen instructions are one line: "No leetcode. No whiteboard. Show us your best AI-built work, a PR, a deployed app, or a repo." The must-haves include a track record building with AI tools, ideally a repo with a CLAUDE.md, at least one shipped end-to-end project a stranger can use, schema and data-flow literacy, and clear written communication. TypeScript fluency from scratch is explicitly not required.
That is a complete hiring rubric pasted into a public job ad. Most teams bury this stuff three rounds deep. AES inverted it.
This is not one weird company
The same June thread has the same pattern from very different employers. Subtle Medical's listing describes a process with a phone screen, a culture-fit interview, a technical interview with a real-world problem (no LeetCode), and a cross-functional team interview. EggAI says it uses AI in both development and product, and screens for engineers who are critical thinkers and understand the limitations of AI but also where it can have positive impact. A trend analysis of the full June 2026 thread found postings referencing AI engineers, agentic development environments, LLM product surfaces, evaluation work, prompt and retrieval experimentation, secure developer infrastructure, and full-stack engineers who can ship with AI in production.
The signal is no longer "we use AI internally, BTW." The signal is "the interview tests whether you can supervise it."
Why LeetCode died and what replaced it
The argument against algorithm interviews is no longer theoretical. AI passes them. DistantJob put it bluntly: while the manual coding era is not entirely dead, AI can solve LeetCode problems in seconds, so testing humans on algorithms no longer measures the ability to build software, it measures memorization or prompt skill.
What replaced the signal is messier and more honest. Google's VP of Engineering Megan K (quoted in Pensero's enterprise review): "AI writes a high volume of code fast, but that code is not inherently production-ready. It is frequently almost right, passing basic tests but containing hidden security flaws, performance regressions, or architectural inconsistencies." That sentence is a job description. The person you want on the team is the one who reads the diff and notices the regression.
HackerRank's own framing of the underlying numbers: role-specific assessments are demonstrably fairer than generic LeetCode-style algorithmic tests, and companies using structured assessments cut false positives in hiring by 50% compared to unstructured screening. The hiring win from killing the algorithm round is not softer interviews. It is a tighter funnel.
The five-item rubric, lifted from AES
Here is the rubric AES essentially publishes, reformatted for use. SurePrompts notes a code-review rubric should be a short checklist of what counts as an issue, with five to seven items the sweet spot and a severity scheme (Critical, High, Medium, Low). The AES JD already lands in that range.
1. Schema literacy
Give the candidate a Prisma schema (or whatever you use) with a subtle modeling bug: a missing unique constraint on a join table, a denormalized field that will drift, a nullable column that should not be. Ask them to read it and tell you what will break in production. This screens for the thing AI consistently gets wrong: data shape over six months of writes.
2. Diff-reading under time pressure
Hand them a 200-line AI-generated PR against a real repo. Ask them to review it like they would on a Tuesday. Score for what they catch (and what they trust). Tools like CodeRabbit, Greptile, Cursor BugBot, and Qodo are now part of the working stack; a candidate who has never disagreed with one of these tools has not done the job yet.
3. End-to-end ownership
Demand a deployed project a stranger can use. Not a repo. A URL. AES's "ideally with a CLAUDE.md" line is the tell: the CLAUDE.md is a portfolio piece now, because it shows how a candidate thinks about constraining an agent against a real codebase.
4. The "wrong code" question
KORE1's six-minute screen, which you can run on a first call: "Walk me through the last time AI gave you code you didn't end up using. What was wrong with it, and how did you figure out it was wrong?" The answer reveals more than any algorithm round. Candidates who have never thrown out AI output are either lying or have never shipped.
5. Written communication on a PR
Pull one of their public PRs. Read the description and the review comments. If they cannot explain a change in two paragraphs to a teammate, they cannot direct an agent either.
That is your screen. Five items, severity-scored, defensible.
The sourcing problem this creates
Here is the catch. If the screen is "show me a PR you directed an AI to produce against a published stack," then the top of the funnel has to look different too. LinkedIn title search will not find these people. "Senior software engineer" matches everyone. The signal is in the artifacts: a public CLAUDE.md, a deployed Next.js 14 RSC app, a thoughtful comment on a CodeRabbit review, a personal repo that uses the Anthropic SDK in production.
That is exactly the search problem Refolk was built for. You describe the engineer in plain English ("ship-credible full-stack devs with a public CLAUDE.md and at least one deployed AI-driven side project") and get a ranked list across GitHub, LinkedIn, and the open web with the artifacts attached. The screening evidence is already in the profile. The first call gets to skip "do you actually use these tools" entirely.
What AES quietly figured out that most teams missed
Five non-obvious wins from this JD. Steal them.
The JD is the rubric. If your posting does not list schema literacy, diff-reading, and end-to-end ownership (or your team's equivalents) in plain language, your funnel fills with 2018-shaped candidates regardless of how the actual screen runs. The JD is upstream of the screen.
"No LeetCode" is not lowering the bar. AES still requires shipped end-to-end work, schema fluency, and a CLAUDE.md. That is harder to fake than memorized binary trees. Removing the whiteboard round raises the resume-to-decision signal, not the opposite.
Stack disclosure is now a screening tool. AES publishes Next.js 14 (App Router, RSC, Server Actions), TypeScript strict, Prisma plus Postgres, Tailwind plus shadcn/ui, NextAuth, Twilio, and the Anthropic SDK. That is not transparency theater. It lets serious candidates produce a stack-relevant artifact before the first call. Publishing the stack is the take-home.
The anti-cheating panic flipped. In 2024 and 2025, the industry was worried about candidates using AI tools to cheat coding interviews. In 2026, AES is asking candidates to show them the AI in the artifact. The CLAUDE.md is not tolerated. It is requested.
Hiring for the wrong skill is now a velocity bug. If AI-generated code volume outpaces human review capacity by 40%, every engineer you hire is, in practice, a senior code reviewer first and an author second. A team optimized for from-scratch authoring will ship slower than one optimized for fast, careful review.
A 30-minute rewrite of your own JD
Open whatever req you have in Greenhouse right now. Do four things.
- Add a sentence naming your codegen layer. "We use Claude Code (or Cursor, or GitHub Copilot Workspace) as our default authoring environment." Candidates self-select.
- Publish the stack. Exact versions. This is the take-home.
- Replace any line about "strong CS fundamentals" with "schema literacy, diff-reading, and end-to-end ownership." If those words feel too narrow, your JD was hiding behind the abstraction.
- Add the screen instruction in the JD itself: "Send a link to a PR or deployed project that shows how you direct an AI in production. A CLAUDE.md is a plus."
Then go find the candidates who fit it. A natural-language sourcing pass through Refolk against "engineers with public CLAUDE.md repos who ship deployed apps on the Anthropic SDK" gets you a working shortlist in the time it took to read this post. The screening artifact comes with the profile.
The trap to avoid
There is a way to do this badly. If you publish the AES-style screen but then run a four-round process that still ends in a live algorithm round "just to be safe," you have done worse than nothing. You have signaled the new bar, attracted the candidates who can hit it, and then filtered them out with the old one. Pick a screen and trust it. AES did.
FAQ
Does removing LeetCode mean we hire weaker engineers?
No, and the data points the other way. HackerRank's own framing is that structured, role-specific assessments cut false positives in hiring by 50% compared to unstructured screens. AES's bar (shipped end-to-end project, public PR, CLAUDE.md, schema fluency) is harder to fake than memorized algorithms. You are trading a noisy signal for a higher-resolution one, not lowering standards.
What is a CLAUDE.md and why does it count as a portfolio piece?
A CLAUDE.md is a repo-level configuration and context file that tells Claude Code (and similar agents) how to behave inside a specific codebase: conventions, where to find the schema, what not to touch, how to run tests. A good one signals that the engineer has thought carefully about constraining an agent against real code, which is the exact skill AES and similar employers are screening for. It is becoming the 2026 equivalent of a well-written README.
How do we source candidates whose signal lives in artifacts, not titles?
Title search is the wrong tool. The signal is in public PRs, deployed projects, CLAUDE.md files, and contribution patterns on AI-tooling repos. That is what Refolk is built to retrieve. You describe what you want in plain English (stack, artifact type, location) and get a ranked list with the artifacts attached, so the first screening pass happens before the first call.
What is the single fastest change a hiring manager can make this week?
Add KORE1's six-minute question to your first call: "Walk me through the last time AI gave you code you didn't end up using. What was wrong with it, and how did you figure out it was wrong?" It costs nothing, replaces no existing round, and gives you a sharper read on AI-fluent engineer screening than most full take-homes. Then go rewrite the JD.