Auto Apply

Software Projects · Auto Apply

n8n · LLM automation · OpenClaw · Oct 2025

Two attempts at the same question: how much of the job application grind can you actually automate? One built on n8n as a data pipeline, one rebuilt as a fully agentic system on OpenClaw.

The problem

People burn absurd amounts of time tailoring resumes, retyping the same information into every portal, and manually scraping job boards. Every step is repetitive. Every form field is a slight variation of the same fifteen questions. And yet we do it all by hand, one application at a time.

I wanted to see how far this could be pushed.

V1, the n8n pipeline

The first version is an n8n workflow. I picked n8n on purpose: I wanted to get comfortable with data flow thinking. How events move between systems, what lives where, how to drop an LLM into a graph as just another node.

entry_id=1_cqlxezlv widget_id=1_s6gjgdy6 title=n8n-resume

Architecture

• Notion as the database for job listings, application state, resume variants, notes, and follow ups
• Scheduled scrapers as n8n triggers, polling job sources for new listings and pushing them into Notion
• Google Drive as permanent storage. Every tailored resume and cover letter lands there with a consistent filename
• LLM nodes inside the graph for rewriting, matching jobs against my profile, and scoring how worth applying a listing actually is

Building this taught me more about data flow between modules than any one off script ever could. When you are forced to model "how does a new job listing become a customized PDF in Drive" as a graph, the shape of the whole system becomes obvious, and you start seeing where the leaks are.

V2, the agentic version (OpenClaw)

Second version rebuilt the same flow as a fully agentic system using OpenClaw. One agent loop plus tool calls instead of a pre wired pipeline.

What OpenClaw is great at: prototyping. I had an end to end version running in a day. Zero plumbing. Just describe the tools, write the goal, let it rip.

Where it falls over: production. Rate limits, slow tool calls, reasoning drifting on long runs, unpredictable cost. For anything repetitive and deterministic, and job applications are extremely repetitive and deterministic, a user script is still more reliable than an agent.

What I am keeping from it: the heartbeat pattern. A persistent agent that survives across long running sessions and keeps its state updated is a genuinely useful primitive. That part I will reuse.

Lesson: agentic frameworks are ideal for MVPs, for validating the idea and proving the loop works. Once the loop is proven, rewrite the hot path as something boring and reliable.

The extension side of the problem

Along the way I ended up benchmarking some off the shelf tools:

• JobRight, a browser extension that auto fills most application fields. Solid, well scoped, pleasant to use. This is the bar.
• Simplify Jobs, a similar idea but not as powerful. It noticeably degrades browser performance by injecting content scripts on every page. I filed feedback with the Simplify team about this.
• Claude in Chrome and OpenClaw browser tools. Both can drive a real browser, but neither is production grade yet. Click fidelity is low, latency is high, and they get confused by dynamic forms.

Takeaways

1. Use a dedicated browser profile for job applications. Do not let extension heavy job automation contaminate your daily profile. The performance hit compounds fast, and you will end up blaming Chrome.
2. Agentic frameworks are great for MVPs and bad for hot paths. Prototype with OpenClaw or n8n's AI nodes, then rewrite the repetitive bits as plain scripts.
3. Notion makes a surprisingly good backend for personal automation projects. Structured enough to query, flexible enough to hand edit when the automation gets it wrong.
4. The bottleneck is no longer the LLM part. Content generation is solved. The bottleneck is the last mile: clicking, filling, navigating ATS systems. That is where the interesting engineering lives now.

What's next

I am focused on the browser level automation piece. Specifically: a userscript that combines the reliability of hand written selectors with the flexibility of an LLM fallback when a form shape is new. Aiming for the productivity of JobRight with the extensibility of an agent.

This is the plain-HTML mirror served to crawlers, LLMs, and curl. Humans with a JavaScript-enabled browser see the rich React/XP-themed SPA at the same URL.

All plain pages · Live site · sitemap.xml