The 0.49% Signal: What a 2026 New-Grad SWE Resume Must Say

A Blind thread titled "Claude just cooked entry level roles and sde's" is sitting on Teamblind's Most Read rail next to "Apple openly hires cheaters" (17,787 views) and "Meta's layoffs are happening and I am scared of my future" (13,010 views). If you are a 2026 CS grad, the thread is not the problem. The problem is that every recruiter screening your resume this week read it too, and now benchmarks your bullets against Claude Code instead of against the kid two seats down from you in 6.172.

This is a tactical rewrite guide, not a pep talk. The CRUD bullets have to go. Here is what replaces them.

Why your "built a REST API" bullet now reads as a liability

Recruiters assume Claude can produce every CRUD-shaped line on an entry-level SWE resume, so leading with that work actively hurts you. Anthropic published the exact numbers hiring managers have internalized: as of May 2026, more than 80% of code merged into Anthropic's own codebase was authored by Claude, and the typical engineer there is now merging 8x as much code per day as in 2024.

The volume problems a junior used to grind through are the ones a model eats first. From an engineering-leads survey this year: "Claude 3.5 and Claude 4 are eating through a very specific type of work, the kind junior developers spent most of their time on in 2022. Boilerplate, CRUD endpoints, test scaffolding, documentation, migration scripts. Not the interesting problems. The volume problems."

Delete the following bullets from your resume today:

• "Built a REST API with Flask and JWT auth"
• "Achieved 95% test coverage on module X"
• "Wrote migration scripts for a Postgres schema change"
• "Developed CRUD endpoints for a to-do list / recipe / booking app"
• "Documented internal APIs using Swagger"

Not because the work is bad. Because the work is now the default output of a $20/month subscription, and your resume is implicitly asking the reader "what do I offer beyond that subscription?"

The scarce signal: 0.49% of new grads will admit they ship with LLMs

The resume move that actually differentiates you in 2026 is proof you shipped with an LLM in a reviewable artifact, because almost no new grad will put it on paper. In Refolk's index of professional profiles, there are roughly 371,952 entry-level Software Engineers and SDEs in the United States. Of those, only 1,836 list LLM or LangChain skills. That is 0.49%.

The gap is not technological. The gap is social. New grads strip AI tooling off their resumes because they think it makes them look "legitimate." It does the opposite. Hiring managers who just read Dario Amodei predicting a one-person billion-dollar company in 2026 are scanning for leverage, not for the pretense that you handwrote every curly brace.

What "shipped with AI in a reviewable artifact" actually looks like on a resume:

1. A public GitHub repo containing your Claude Code or Cursor session transcripts, with your own commentary on where the model was wrong.
2. An agent eval script. Not a demo, an eval: inputs, expected outputs, a scoring function, a holdout set, and a number.
3. PR descriptions in your own voice explaining what the model proposed, what you rejected, and why.
4. A measurable production outcome where the LLM was a tool, not the headline. "Triaged 400 GitHub issues with a 3-agent pipeline, 92% precision on a 50-issue manual holdout" is a bullet. "Used ChatGPT" is not.

Keep the on-call. Delete the CRUD.

The work that justifies hiring a human junior in 2026 is the work a model cannot take responsibility for at 3 a.m., so your resume has to lead with evidence of exactly that. The top-voted reply under the viral Blind thread said it cleanly: "claude is basically a cracked intern but it doesn't understand a decade-old codebase or take responsibility when prod blows up at 3am. writing code was never the hardest part anyway."

A separate Blind thread this week, "PE told me, once operations get partially automated, SDE headcount is cooked," landed the mechanism: "The plan across a number of orgs is to use AI to automate ops investigations and ticket handling. A lot of SDE headcount is actually just meant to keep a rotation for brutal oncall." On-call rotation was the pipeline new grads walked into. If your resume shows you can take a page, that pipeline is still open to you.

Bullets to add if you have the receipts:

• "Carried primary on-call for a 4-service surface area over a 12-week internship. Wrote 7 postmortems; 2 drove permanent fixes."
• "Debugged a cross-system latency regression spanning the mobile client, a Kong gateway, and a downstream Rust service. Root cause: a retry budget misconfig."
• "Rewrote the runbook for X incident class. Mean time to acknowledge dropped from 14 to 4 minutes over the next quarter."
• "Owned the migration of logging from stdout scraping to OpenTelemetry across 3 services. Shipped behind a flag, backed out once, re-shipped clean."

None of those are things a model does unattended. All of them are things a hiring manager pattern-matches to "this person will survive their first real outage."

Tailoring these bullets to each specific posting is the exact work Refolk takes off you: paste the job description, get your own resume back rewritten to lead with the artifacts that posting actually screens for, with the CRUD lines demoted or killed.

"Orchestrator" is the real 2026 SDE1 job description

The job title on the req still says Software Engineer, but the work underneath it is orchestration, so your resume should use the verb "orchestrated" with LLM agents as the object. From the same engineering-leads survey: "what's really happening is a shift from writing boilerplate to acting more like orchestrators."

The frontier version of this looks like Nicholas Carlini's experiment, where 16 Claude Opus 4.6 agents wrote a C compiler in Rust from scratch, capable of compiling the Linux kernel, at a cost of roughly $20,000 in API calls. Your version is smaller, cheaper, and shaped the same.

A new-grad orchestration bullet, broken down:

• Verb: orchestrated, scheduled, routed, evaluated.
• Object: a named model or agent framework (Claude 4, GPT-5, LangGraph, DSPy, Inngest).
• Scope: how many items, over how long, against what source of truth.
• Verification: your own eval, with a number and a holdout.
• Cost: dollars per run or latency per task, because leverage is the point.

Example: "Orchestrated a 3-agent pipeline (planner, retriever, writer) over Anthropic's Claude 4 and a local 7B reranker. Processed 4,200 support tickets in 11 days at $0.021/ticket. Verified on a 200-ticket holdout: 89% category accuracy, 94% PII redaction recall."

That bullet survives a 20-second screen. "Built a chatbot" does not.

The Austin case, and what it means for your target list

A mid-sized Austin SaaS company cut its junior engineering headcount by 40% this year by not re-hiring contractors, with Claude Code cited by name as the reason. The quiet layoffs are concentrated at companies that treated SDE1s as capacity, not as apprentices. Your target list should invert that.

In Refolk's index, the employers currently absorbing the most entry-level SWEs in the US include Google, Figma, Microsoft, LinkedIn, Ashby, and Glean. Note the last two. Glean and Ashby are AI-native companies hiring juniors who ship with LLMs natively, which means your "orchestrated a 3-agent pipeline" bullet is a feature at those shops, not a curiosity.

How to triage your application list this week:

1. AI-native companies hiring juniors. Ashby, Glean, and the smaller Series B shops where the founding engineers already use Claude Code daily. Your AI bullets are table stakes, which is good; it means they expect them and will read them carefully.
2. Big-tech new-grad programs with structured on-call rotations. Google, Microsoft, LinkedIn still run the apprenticeship model. Lead with on-call, debugging narratives, and the one orchestration project.
3. Avoid "junior as capacity" shops. Agencies, consultancies, body shops, and any posting that reads like a scope-of-work. Those are the roles getting Claude-replaced first.

What the "About" or summary line should actually say

Your top-of-resume summary should assert learning velocity, not current output, because that is the only story that still justifies hiring a human junior in 2026. The highest-signal comment under the Blind thread put it plainly: "You don't hire a junior for the shitty code they can write today, you hire them because they'll learn to write better code tomorrow and be a senior engineer some day."

Translate that into resume language. Signals that read as learning velocity:

• A commit graph that shows a measurable skill jump in 90 days. Link it.
• A reading log or "what I shipped this quarter" page on a personal site, dated.
• Two or three incrementally harder projects, in order, with the second one visibly correcting mistakes from the first.
• A line like: "Shipped 4 production features in a 12-week internship; by week 10, PRs required zero correctness revisions from my reviewer."

Signals that read as static output, and therefore dated:

• "Proficient in Java, Python, C++, JavaScript, TypeScript, Go, Rust, SQL."
• "Strong problem-solving skills."
• "Passionate about technology."
• Project lists with no dates.

The mechanical work of pulling the right projects out of your GitHub, writing them as outcome-shaped bullets, and matching them to each specific posting is what Refolk automates for new grads. Your history goes in; a resume tailored to the posting in front of you comes out, with a fit score telling you whether to bother applying at all.

The cheater foil: do not be the Cluely story

Recruiters are actively pattern-matching for undisclosed AI use in interviews, so disclose everything on the resume and over-document the artifact. Cluely and Interview Coder 2.0 trended on Blind the same week as the "Claude cooked SDEs" thread, with "Apple openly hires cheaters" as the sibling post. Being the orchestrating-agents-in-production candidate is a resume asset. Being the cheating-in-interviews candidate is a background-check risk.

The distinction on your resume is simple:

• Disclose the model by name in your bullets.
• Show your eval, with a holdout you built.
• Describe what you rejected from the model, not just what you accepted.
• Keep the artifact public and the commits dated.

That is the shape of a 2026 new-grad SWE resume that survives the Claude benchmark. Everything else is 2022.

FAQ

Should I remove AI projects from my resume if I am applying to a traditional big-tech new-grad program?

No, but reframe them. For Google, Microsoft, or LinkedIn new-grad reqs, keep the orchestration project but position it as a systems project with an LLM in it, not as an "AI project." Lead the bullet with the system (ingestion pipeline, triage service, eval harness) and the measured outcome, then name the model. Those programs still hire for CS fundamentals and on-call survivability; your AI work demonstrates leverage without displacing the fundamentals they screen on.

What if I only have CRUD-shaped internship work? I literally built a Flask API.

Rewrite it around what broke, what you measured, and what you decided, not what you built. "Built a Flask API with JWT auth" becomes "Owned auth for a 4-endpoint service used by 60 internal users; debugged a token-refresh race that caused 3% of sessions to drop, shipped a fix behind a flag." Same project, different signal. The reader now sees ownership and debugging, which are the parts a model cannot do for them.

Is it worth listing LangChain, LlamaIndex, or DSPy if I only used them in side projects?

Yes, as long as the side project has an eval and a number. The 0.49% figure from Refolk's index means almost nobody at your level lists these at all, so even a modest, well-measured side project puts you ahead of 99% of the cohort on paper. The failure mode is listing the framework with no artifact; recruiters have learned to ask about it. If you can produce a repo, an eval script, and a two-sentence story about a failure mode you caught, the bullet earns its spot.

How much of my resume should be AI work versus traditional SWE work?

Roughly one orchestration project, one on-call or debugging narrative, and one systems project, in that order, with the rest of the page carrying your coursework, internships, and the usual skills block. The point is not to become an "AI engineer"; the point is to show the reader that one of your three headline projects answers the "why not just use Claude" objection before they ask it. Three bullets is enough to clear the bar. Five is overfitting.