One hour. One workflow. Everyone ships.

Most of the hour is hands-on. We'll build a real CV-screening workflow together, node by node. The first 5 minutes set the stage; the last 10 are for questions and where to take this in the hackathon.

n8n is a visual builder for the work between your tools.

You drag nodes onto a canvas, wire them together, and the workflow runs. Each node is either a trigger (something happens), an integration (talk to Slack / Sheets / GitHub / 400 others), or a helper (transform data, branch on conditions, loop). AI nodes are first-class - built right into the same canvas.

The five things to do and the five to avoid.

Indeed's internal guidelines are explicit about how n8n is meant to be used. Before you build, read this. Especially the credential-hygiene and AI-policy lines. Full AI policy: indeedhuddle.com/huddle/company-wide-generative-ai-use-guidelines.

Code is #1 and that's a problem.

Top 10 nodes used across Indeed's n8n platform. The Code node is the most-used node and also the most-misused - most of the time, a purpose-built node already does the job. Reach for Set, IF, or a native integration first.

Form → AI → CSV. Eight nodes. One canvas.

A recruiter pastes a job description and uploads a stack of CVs. The workflow extracts each CV's text, sends it to GPT with the JD, gets back a typed JSON object per candidate, and assembles a downloadable CSV with scores and one-line reasons.

Three things, then we're building.

If you came with the pre-workshop checklist done, you can skip ahead. If not, the next slide has the JSON for download - work through these in order.

Grab the workflow and the sample data.

Two downloads. The workflow JSON imports into n8n. The sample data zip has a Senior Frontend Developer JD (copy-paste into the form) and 5 fake CVs (upload all at once).

The front door of the workflow.

n8n's built-in form. Hosted at a public URL (or behind your network). Two fields: a Job Description textarea and a multi-file CV upload. The responseMode: lastNode setting keeps the user's session open until the very last node returns the CSV download.

Why a Form (not a Webhook)? Forms give you a usable UI for free. For the hackathon, swap to a Webhook to receive CVs from your careers portal, or an Email Trigger to auto-process applications.

Configuration

Form TitleCV Screening

Field 1Job Description (textarea)

Field 2CVs (file, multiple)

responseModelastNode

Both fieldsrequired: true

Normalize Files. Yes, this one is Code and that's OK.

Multi-file uploads come in as CVs_0, CVs_1… on the same item's binary payload. We need one item per CV downstream. There's no native node for binary fan-out, so this is the doc's allowed Code use: "logic that no built-in node covers."

Lesson: the doc says Code is the #1 yellow flag - but it's not banned. The test is "is there a Set, IF, or native integration that already does this?" For binary fan-out, the answer is no. Keep it.

What it does

Input1 item · N binaries

OutputN items · 1 binary each

Alsocopies JD onto each item

TypeJavaScript (12 lines)

The pattern most beginners get stuck on.

SplitInBatches processes one CV at a time. It has two output branches: the loop branch (off to do the work, then comes back) and the done branch (runs once when all CVs are processed). The wire that goes back into the node is the loop - not a mistake, it's how batch processing looks in n8n.

If you only remember one thing here: branch 1 is the loop body, branch 2 is what runs after the loop finishes. The back-edge wire is intentional.

Configuration

Batch Size1

Output 1 (loop)→ Extract PDF Text

Output 2 (done)→ Convert to CSV

Loop back fromFlatten Output

Reads every page into a single string. Native node, no setup.

n8n's built-in Extract from File node, configured for PDF mode. It pulls the text out of every page and joins them. No external API call, no credential - just text in, text out.

Hackathon upgrade: for scanned CVs, complex multi-column layouts, or non-English text, swap this for the native Mistral AI node (OCR operation). Higher quality, tiny code change, much better downstream AI results.

Configuration

OperationPDF

Binary propertycv

Join pagestrue

Output field$json.text

An AI Agent with a typed schema.

This is the magic. The Agent node sends the CV text + the job description to GPT with a system message ("you're an expert HR recruiter"). The Structured Output Parser sub-node forces the model to return a strict JSON object - every field typed, no string-parsing fragility, no markdown fences to strip.

Three sub-nodes hang off the Agent: the OpenAI Chat Model (which model, with your credential), the Structured Output Parser (the schema), and an optional auto-fix model for when the LLM occasionally drifts off-spec.

Configuration

Modelgpt-5-mini

System message"Expert HR recruiter…"

Output schema9 fields, typed

Schema fieldsname, email, role, score, …

hasOutputParsertrue

Unwrap the AI output into clean CSV columns.

The Agent's output comes wrapped in $json.output.*. The Set node unwraps it into nine top-level fields - one expression per output column, with the skills array joined into a string. No JavaScript, readable by anyone on the team.

Pattern worth keeping: for mapping fields, joining arrays, or applying defaults, reach for Set first. One expression per field, no JavaScript, anyone on the team can audit it.

Field mappings

fileName{{ $json.output.fileName }}

fullName{{ $json.output.fullName }}

topSkills.join(", ")

score{{ …score }} (number)

… 5 moreall from $json.output

Turn the rows into a downloadable file.

Native Convert to File node in CSV mode. Takes every flattened row from the loop and produces a binary file named cv_screening_results.csv with a header row generated from the field names.

Why a CSV? Universal, opens in Excel / Sheets / Numbers, easy to share. For a hackathon project you'll want to swap this for something else - see the next slide.

Configuration

OperationCSV

fileNamecv_screening_results.csv

headerRowtrue

Outputbinary file

Hand the file back to the browser tab that started it.

The Form node in completion mode with respondWith: returnBinary. The user's browser tab - still open from the original form submission - gets a "Screening complete!" page with a download link to the CSV produced in the previous step.

Hackathon upgrade: swap this exit for an ATS push, a Sheets / Notion row append, or a Slack DM to the recruiter with the top-3 candidates. The exit step is the easiest thing to change.

Configuration

operationcompletion

respondWithreturnBinary

completionTitleScreening complete!

inputDataFieldNamedata

CV scoring is regulated. Treat it that way - everywhere.

Today's workflow is a learning demo. Real CV screening at Indeed needs HR sign-off, legal review, and a region-specific compliance check. The rules differ by jurisdiction; the underlying principles - bias auditing, candidate transparency, human oversight - don't.

Five places this workflow can go next.

Today you built the spine. The hackathon is for putting muscle on it. Each of these is a half-day project - pick one, fork the workflow, ship it.