AES Just Told HN "The Bar Is NOT Coding." Screen for CLAUDE.md Instead.
A June 2026 HN job post deprioritized coding entirely. Here's how to source and screen the AI-director engineer archetype without breaking your funnel.
A June 2026 "Ask HN: Who is hiring?" post from AES, an environmental-test-chamber manufacturer in Chelmsford, said something most JDs still won't: "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." If your screener still opens with a LeetCode link, you are now measuring the wrong axis for a growing class of teams.
This isn't a niche opinion from one small shop. It's the leading edge of a screener redesign happening at Meta, Google, Canva, and Sierra. The recruiters who see it early get the archetype before it has a name on LinkedIn.
What AES actually said, and why it matters
The exact language in the post: "The senior lead doesn't write code from scratch. Neither will you." The role is a summer software engineering intern, remote US, $30 to $40 per hour, June 15 to August 22, 2026. Stack is Next.js 14 with App Router and Server Components, TypeScript strict, Prisma and Postgres, Tailwind and shadcn/ui, NextAuth, Twilio, Anthropic SDK. Team shape: one senior lead, one part-time backend specialist, one intern.
Read the requirements carefully. TypeScript fluency is explicitly not required from scratch. What is required: "a track record building with AI tools, at least one shipped end-to-end project that runs somewhere a stranger can use, schema and data-flow literacy, clear written communication." The process line: "No leetcode. No whiteboard. Show us your best AI-built work, a PR, a deployed app, or a repo, ideally with a CLAUDE.md."
That last clause is the sourcing tell. AES is not asking for a self-declared skill. They are asking for a file.
AES is not alone in the thread
Memora, a five-person EU startup, wrote: "We're AI-native, but not trying to replace human creativity. We use tools like Cursor and Claude Code as leverage." Foam, in SF, is paying $300 to $400k for a staff founding engineer expected to have "an understanding of the internals of things like Claude Code, and a broader understanding of orchestrating agents to complete complex discrete tasks." Comvex gates its customer success role on "100+ hours of real hands-on experience building AI agents and automations. This is a genuine bar, if you haven't, this isn't the role."
Three JDs, three functions (intern, staff eng, CSM), all triangulating the same competency: direct the model, verify the model, own the outcome.
Why the old screener misses this archetype
The market shifted underneath conventional coding tests. Stack Overflow 2025 puts daily AI-tool use at 51% of professional developers. DORA respondents spend a median two hours per day on AI-assisted work. The 2026 numbers push further.
Here's the workflow flip that justifies AES's position. Developers now spend 11.4 hours per week reviewing AI-generated code versus 9.8 hours writing new code. That is a reversal of the 2024 pattern. The job has already changed from author to editor for most teams, but the interview loop hasn't followed.
The frustration data explains why "catching it when it's wrong" is a real competency, not a slogan. 66% of developers say the top pain point is "AI solutions that are almost right, but not quite." 45% say debugging AI-generated code takes more time than writing it manually. And the velocity story from the same research: pull requests per developer rose 20% with AI assistance, but incidents per PR rose 23.5%. More output, more breakage. The marginal engineer is worth less for how much they can ship and more for how much of what the model shipped is actually right.
The job flipped from author to editor. Most interview loops still test for authorship.
The sourcing problem: nobody self-labels yet
Here is the trap. A search across public US professional profiles for people who explicitly feature "Claude Code" in their headline or skills returns fewer than 30 people. The top employers of that tiny cohort are consulting shops (Cognizant, BitWise, SAK, Climb Global), not product engineering teams. The archetype exists. It doesn't self-label.
If you type "Claude Code" into LinkedIn Recruiter and stop, you will conclude the talent pool is thirty consultants. That is wrong by three orders of magnitude.
The signal has moved to file names in public repos. AGENTS.md is becoming the cross-tool convention. CLAUDE.md is Claude Code specific. Both encode project rules, conventions, and operational expectations so agents can reuse them without restating context every session. The presence, structure, and specificity of a candidate's CLAUDE.md is a higher-signal proxy for AI-director skill than any bootcamp cert, skill tag, or coding-test score.
This is why keyword sourcing is broken for this role and why we built Refolk: you describe the archetype in plain English ("engineers with a CLAUDE.md or AGENTS.md in a public repo, shipped a Next.js App Router project, based in the US") and get a ranked shortlist across GitHub, LinkedIn, and the open web. The self-declared skill tag is the wrong index. The artifact is the right one.
Practical GitHub queries that beat LinkedIn Boolean
Start with the file. filename:CLAUDE.md and filename:AGENTS.md are both indexable. Filter by recent commit activity, stars, and by whether the enclosing repo actually deploys somewhere. A CLAUDE.md in a dead repo is a form of AI theater. A CLAUDE.md in a repo with a live Vercel URL, real users, and non-trivial Prisma migrations is a signal.
Then read the file. A weak CLAUDE.md is one paragraph of "please use TypeScript." A strong one names schema boundaries, migration procedures, testing gates, error-handling conventions, and where the model is explicitly not allowed to touch. That gradient is the interview.
The screener redesign: what Meta, Google, Canva, and Sierra are already doing
The Fortune 100 is catching up to what AES already put in the JD, six to twelve months behind. Meta rolled out an AI-enabled coding interview in October 2025 that replaces one of two traditional coding rounds at the onsite stage. Google is following with a round that "includes reading, debugging, and optimizing real code with Gemini available as an AI assistant. Interviewers will evaluate AI fluency, including prompt engineering, output validation, and debugging skills." Google's internal doc describes the format as "human-led, AI-assisted."
Canva went further and made it a competency, not a bonus round. In their words: "We've piloted a new competency we called AI-Assisted Coding that replaces our traditional Computer Science Fundamentals screening for backend and frontend engineering roles."
Sierra is running the format that maps almost one-to-one to what AES is describing: a debugging interview where candidates are given a medium-sized codebase and a draft PR from a colleague that introduces a cross-cutting feature. The candidate's job is to review and improve it, pulling down the code, inspecting the output, and iterating with coding agents to make it better.
Notice the common thread. None of these companies decided "coding is dead." All of them decided the evaluation context changed. Sierra and Canva both say directly: "code fluency and technical depth were still absolute requirements, just evaluated in a different context."
What to actually put in the loop
Four rounds that reflect the AES bar without needing to be AES:
-
Artifact review. Candidate walks you through a public repo they built with heavy AI assist. Ask them to show the CLAUDE.md, explain what they chose to constrain, and describe one time the model got it wrong. This replaces the resume walkthrough.
-
PR triage. Give them a real draft PR from your codebase with two or three subtle defects (a Prisma migration that's non-reversible, a Server Action missing auth, a schema field that breaks a downstream query). Watch them use whatever coding agent they prefer. Score on catches per minute and on the questions they ask before touching the diff.
-
Schema and data-flow. Whiteboard-adjacent, but the whiteboard is a data model. RSC boundaries, Server Actions, cache invalidation, migration safety. This is where AES's "raises the bar on schema" actually gets tested.
-
End-to-end ownership scenario. "The model shipped a feature that passed CI. Production error rate is up 4% in the last hour. Walk me through what you do." This tests instinct for the velocity-vs-stability paradox.
None of these require a LeetCode subscription. All of them are hard to fake.
Do not confuse AI-native with junior or cheap
The most dangerous misread of the AES post is "AI does the coding, so I can hire cheap interns." Foam is paying $300 to $400k for a founding engineer expected to understand Claude Code internals. Comvex gates a non-engineering role on 100+ hours of agent building. AES itself is paying its intern $30 to $40 per hour and expecting a fully shipped end-to-end project as an application artifact.
This archetype is senior. It is disguised as junior-friendly because the syntactic barrier is lower. Founders who read the JD as permission to under-hire will build unmaintainable products, and the incident-per-PR data (up 23.5%) predicts exactly how that fails.
If you are staffing an AI-native team, the sourcing brief is: senior judgment, mid-level compensation flexibility, portfolio-first evidence. Boolean searches on LinkedIn Recruiter cannot find this cohort at scale because the cohort is defined by artifacts, not tags. Describe the archetype to Refolk in plain English and it will pull from GitHub file contents, deployed project URLs, and open-web signals in one pass. That is how the AES intern applicant pool gets found without a $30k Recruiter seat and six weeks of Boolean iteration.
The 90-day window for recruiters
Tracking is easy if you know where to look. hnhiring.com and aitmpl.com/jobs both aggregate JDs that explicitly require Claude Code or agentic tooling. The month-over-month growth in JDs using AES's exact framing ("we use Claude Code, the bar is directing it") is the leading indicator. Once that language lands in a Series B JD from a company with 100 engineers, it will be a Meta round within a year.
The recruiting orgs that win the next 90 days are the ones that (a) stop sourcing "Claude Code" as a keyword, (b) start sourcing CLAUDE.md and AGENTS.md as artifacts, and (c) redesign the loop to test PR triage instead of pure authorship. The JDs are here. The candidate self-labels are not. The gap between those two is where a well-tuned sourcing pipeline compounds.
FAQ
How do I source engineers who use Claude Code if they don't say so on LinkedIn?
Search their public work, not their profile. GitHub queries for filename:CLAUDE.md and filename:AGENTS.md return the artifact directly, and the quality of the file itself is the sourcing gradient. Cross-reference with recent commit activity, a live deployment URL, and a non-trivial schema. Refolk lets you describe the artifact and archetype in plain English rather than stitching GitHub search, LinkedIn filters, and open-web scraping by hand.
Does "the bar is not coding" mean I should stop testing code entirely?
No. It means you should stop testing authorship as the primary axis. Sierra and Canva both keep code fluency as an absolute requirement but evaluate it inside AI-assisted contexts, PR review, debugging, and schema reasoning. Meta and Google have replaced only one of two traditional coding rounds. AES-style teams have gone further, but even they demand schema literacy and diff-reading, which are code-native competencies.
What is a strong CLAUDE.md versus a weak one?
A weak CLAUDE.md is a paragraph of generic guidance ("use TypeScript, be careful"). A strong one names specific schema boundaries the model must not cross, migration procedures, testing gates, error-handling conventions, secret-handling rules, and the exact style of PR the human wants to review. If the file reads like an onboarding doc for a new hire, the candidate has internalized the AI-director job. If it reads like a README, they have not.
How is this different from the older "AI engineer" hiring wave?
The older AI engineer role hired people to build LLM features. The AI-director archetype hires people to build any feature using LLMs as their primary implementation tool. AES is not building AI. They are building a field-service platform. Claude Code is the compiler, not the product. That distinction changes both the sourcing brief and the screener design, and it is why the June 2026 HN thread reads differently from any prior year.