AES Just Posted a $30/hr Intern JD That Kills the Leetcode Screen
The June 2026 HN "Who is hiring" thread surfaced an intern JD that replaces leetcode with a CLAUDE.md. Here's how to rewrite your screen.
If you read the June 2026 HN "Who is hiring" thread the way recruiters usually read it (scanning for stack keywords and comp bands) you missed the actual news. A small environmental-test-chamber company in Chelmsford, MA posted a summer intern JD that quietly rewrites what an engineering hire is for. The bar is no longer "can you code." It is "can you direct an AI at production code and catch it when it's wrong."
The post comes from AES (associatedenvironmentalsystems.com). It pays $30 to $40 per hour for a remote summer intern, June 15 through August 22. Applications closed June 8. The team is a senior lead, a part-time backend engineer, and the intern. The stack is Next.js 14 with App Router and RSC, TypeScript strict, Prisma on Postgres, Tailwind with shadcn/ui, NextAuth, Twilio, and the Anthropic SDK. None of that is what made the post travel.
This did, verbatim from the JD:
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.
Then, in the same post: "No leetcode. No whiteboard. Show us your best AI-built work, a PR, a deployed app, or a repo." Must-haves include a CLAUDE.md in the submitted work, one shipped end-to-end project a stranger can use, and what the JD calls "schema/data-flow literacy."
That is the first widely-circulated JD I have seen that treats a prompt and context-engineering artifact (the CLAUDE.md) as a screening deliverable. If you run engineering hiring, this matters more than any layoff list this week.
The triad that replaces leetcode
AES names three things it actually screens for: schema literacy, diff reading, and end-to-end ownership. None of them require typing code from a blank file. All three are measurable in a take-home that takes under three hours.
Schema literacy. Hand the candidate a Prisma schema with a deliberately broken relation (a missing index on a foreign key, a denormalized field that should be a join table, a nullable column the API treats as required). Ask them to explain what will break in production and propose a migration. You learn in 20 minutes whether they understand the shape of data or just the syntax around it.
Diff reading. Paste a 300 to 600 line PR generated by Claude Code or Cursor against a real repo. Ask the candidate to leave review comments. The signal is not whether they catch the obvious bug. It is whether they catch the "almost right" issues: the silently-swallowed error, the off-by-one on a paginated query, the auth check that runs after the data fetch. This is the same skill 66% of developers in the Infobip survey said is their biggest frustration with AI output, that "AI solutions that are almost right, but not quite." You are literally hiring for the gap.
End-to-end ownership. Did they ship one thing a stranger can use? Not a tutorial clone. Not a half-finished portfolio site. One deployed app with a domain, auth, and a database. This filters harder than any algorithmic round because it cannot be crammed for in a weekend.
Why this is not just one weird intern post
The AES JD reads like a one-off until you scan the rest of the same June 2026 thread. The pattern is everywhere.
Cora AI, hiring a Founding Full Stack / Applied AI Engineer in SF or LA at $190 to $250k, writes that candidates must "build with AI IDEs (Cursor, Claude Code) as a force multiplier, not occasionally, as your default." Deep Core Technology, a geologic-modeling shop in Canada hiring a founding engineer at $105 to $125k CAD, requires "an understanding of the internals of things like Claude Code, and a broader understanding of orchestrating agents to complete complex discrete tasks." A separate founding-engineer post in the same thread notes the company "Uses AI to generate 90%+ of their code" and has replaced leetcode with a short paid work trial. Babou.ai puts it bluntest: "agents are how you ship code, and you know when an LLM is wrong. You move fast without shipping slop."
Even a non-engineering role in the thread (a Brazil-based CSM at an AI agent company) bars candidates without "100+ hours of real hands-on experience building AI agents/automations."
This is not a vibe. It is a category shift in what an engineering hire is for. And it is showing up at the intern wage band, the founding band, and the customer-facing band in the same week.
The trust gap is the job
The interesting macro number is not adoption. It is the trust collapse underneath it. Stack Overflow's 2025 survey (n over 49,000) found 84% of developers use or plan to use AI coding tools, but only 29% trust the output, down from 40% in 2024. Laura Tacho's DX keynote, drawing on 121,000 developers across 450-plus companies, put weekly AI-assistant use at roughly 75% and monthly use at 92.6%.
Put those numbers next to the 26.9% of production code that is now AI-authored and you have the JD's logic in one line: the bottleneck role is the human who catches the wrongness. AES is not replacing the intern with Claude. AES is using Claude to make a one-intern team viable that previously was not. That is the contrarian read on what AI is doing to entry-level hiring at small companies. It is enabling it, not killing it, but only for candidates who can do the triad.
If your inbound funnel was tuned for the old bar (CS degree, leetcode medium, two internships at name-brand companies) you will miss most of the people the AES JD is actually written for. They are likelier to have a deployed Next.js side project, a CLAUDE.md in a public repo, and a thoughtful PR review history than a 1600 Codeforces rating. Sourcing them by Boolean on LinkedIn does not work. Sourcing them by what they have built does, which is why we built Refolk: describe the candidate profile in plain English and get a ranked shortlist drawn from GitHub, LinkedIn, and the open web, including the public artifacts (repos, AGENTS.md files, deployed apps) that the new JD actually grades on.
Rewriting your JD this week
Here is the minimum set of changes to bring your req in line with the AES pattern. Treat this as a checklist.
1. Strike the algorithmic language
Delete "strong algorithms and data structures fundamentals" if you mean it as a screening signal. Keep it only if your product genuinely requires it (compilers, databases, low-level systems). Most CRUD-ish product teams do not. Replace with: "comfort reading and reviewing AI-generated diffs across a multi-file change."
2. Name the tools
The AES JD names Claude Code. Cora AI names Cursor and Claude Code. Be specific. The DX-cited stat that the best 2026 developers use 2.3 AI tools on average (typically an AI IDE plus a terminal agent) tells you the shape: ask for fluency with one IDE-integrated tool and one agentic CLI. Vague "experience with AI tools" gets you vague candidates.
3. Make the take-home about review, not authorship
Replace the "build a small app" take-home with: here is a real PR our team generated last week with Claude Code, leave review comments. Two hours, max. Pay for it. You will learn more about how the candidate thinks in their inline comments than in any system-design round.
4. Ask for a CLAUDE.md or AGENTS.md as a deliverable
This is the move I expect to spread fastest. A CLAUDE.md (or AGENTS.md, the GitHub convention) is the candidate's documented theory of how to direct an agent at their specific codebase. It reveals context-engineering taste in a way no resume bullet can. Asking for one as part of the application is the cleanest screening artifact I have seen in a decade.
5. Reframe years-of-experience bands
The METR longitudinal study is the underrated datapoint here. In early 2025, experienced developers using AI were 19% slower. In early 2026, the same developers were 18% faster. Same humans, same tools improving, but also a learned technique. Screening for AI-native engineering JD fit means screening for hours spent learning that technique, not years spent before AI existed. A two-year-out engineer with 1,000 hours of Cursor and Claude Code time will outperform a ten-year-out engineer with 50 hours, on this kind of team.
What this means for sourcing
Three operational shifts for anyone running a pipeline against this new bar.
Public artifacts beat resumes. A CLAUDE.md in a starred repo is a stronger signal than a FAANG line item. Shift your sourcing surface from LinkedIn-first to GitHub-first for at least half your pipeline. Tools that can search across GitHub for specific files (AGENTS.md, CLAUDE.md, cursor rules) and rank by recency and engagement are now the high-leverage sourcing surface. Refolk does this in plain English so you do not have to maintain a Boolean string that breaks every time GitHub changes its search syntax.
The "almost right" reviewer is the rare profile. Look for candidates with a track record of substantive code review comments on open-source PRs, especially on AI-assisted PRs. This is harder to find than it sounds and is exactly the kind of query that does not work as a keyword search.
Stop filtering out non-CS-degree candidates. The AES JD does not require a degree. Neither do most of the JDs in the same thread. Candidates who built shipped things with AI tools rather than grinding LeetCode now have a more legible signal than they did 18 months ago. Your filters should reflect that, or you will pay a premium for credential-matched candidates while your competitors hire from a wider pool.
The June 2026 thread will not be the last signal of this shift. It is the cleanest one yet. If your JD, your screen, and your sourcing stack still grade for "can you code from scratch," you are recruiting against a bar your competitors already moved.
FAQ
Is "no leetcode" really viable for senior roles, or is this just an intern thing?
The same June 2026 HN thread has founding-engineer roles at $190 to $250k (Cora AI) and $105 to $125k CAD (Deep Core Technology) using identical framing. A separate founding role in the thread says the company generates 90%-plus of code with AI and uses a short paid work trial instead of algorithm rounds. The pattern is consistent from intern through founding. The constraint is not seniority. It is whether your product depends on raw algorithmic depth.
What exactly should a CLAUDE.md contain to be a useful screening artifact?
At minimum: a project overview, the data model and key invariants, the conventions the agent should follow (naming, error handling, testing), and the boundaries (files or behaviors the agent should not touch without confirmation). The best ones also document past failure modes ("the agent tends to swallow Postgres connection errors here, always check"). Reading a candidate's CLAUDE.md tells you how they think about context, blast radius, and supervision, which is the actual job.
How do I source candidates whose strongest signal is a public CLAUDE.md or shipped side project?
Stop with the Boolean string. The signal lives in repo contents, deployed apps, and review history, not job titles. You want a sourcing surface that reads across GitHub, LinkedIn, and the open web at the same time and ranks on artifacts, not keywords. That is what Refolk does: you describe the profile in plain English (for example, "engineers with a public CLAUDE.md, a deployed Next.js project, and recent PR review activity") and get a ranked shortlist.
Does this kill the entry-level engineering hire or save it?
The Forrester read was that AI would gut entry-level hiring. The AES JD suggests the opposite at small teams. AI is what makes a one-intern team viable. The catch is that the intern has to clear the new triad (schema literacy, diff reading, end-to-end ownership) on day one, because there is no senior engineer with capacity to teach those skills from scratch. So entry-level hiring survives, but it gets selective in a different way: less credentialed, more demonstrated, and faster to evaluate when the artifacts are public.