One hour. Three checkpoints. Everyone ships.

Most of the hour is hands-on. You build along with me, on a clean canvas, one checkpoint at a time. Each checkpoint also ships as a downloadable JSON so you can catch up if you fall behind. The first 5 minutes set the stage; the last 10 are open Q&A.

Quick demo: the finished workflow.

A manager receives 13 fake job postings in a Google Sheet. The workflow loops over each row, asks an LLM to review it (salary realism, inclusive language, JD improvements), writes the AI's note back to the sheet, and DMs you on Slack with a one-line summary when the whole batch is done. Watch this run end-to-end. Then we wipe the canvas and build it ourselves, three checkpoints at a time.

Three things to do, three to avoid.

Indeed's internal guidelines apply to today's training too. Especially the credential-hygiene line.

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

The name comes from NodeMation: "n" + 8 letters + "n". Drag nodes onto a canvas, wire them together, the workflow runs. Open-source / fair-code, self-hosted at Indeed.

Anatomy of a node.

Every node looks the same. A title and an icon up top. An input dot on the left, an output dot on the right. Click the node to open its settings panel: inputs on the left, configuration in the middle, outputs on the right. Press Escape to close.

On the canvas

Open mode · inputs / settings / outputs

Clone the sample jobs sheet to your Drive.

1. Open the sample sheet using the link below.
2. Click File in the top-left.
3. Choose Make a copy.
4. Name it [Your name] Job Reviewer and save to My Drive.
5. Confirm the AI Note column already exists in your copy.

Sample sheet: Sample Data for Pre-Hackathon n8n session

Don't edit the original, it is shared. Always work in your copy.

← →

Open n8n in the right project.

1. In n8n, switch to the Hackathon 2026 Training project (top-left).
2. Click Create workflow from inside that project (not Personal).
3. Name your workflow using the format Full Name - Jobs Evaluator Workflow.
4. Add the hackathon-prep tag at the top of the canvas, so the platform team can clean up after.

Example Workflows project is auto-shared with everyone on first login. Browse it after the session: Prompt Registry, App Integrations, AI Agents and LLMs, MCP, and more.

Credentials available: Hackathon 2026 OpenAI account 20 and Hackathon 2026 Slack Credentials. Both are pre-configured for this session only.

← →

Five nodes. One full loop iteration.

The placeholder we write back uses an expression that embeds the job title. So every row in the sheet gets a different note, which proves the loop is iterating. Same expression syntax we will use for the LLM prompt in Checkpoint 2, but here you learn it for free.

CP1 build, screen by screen.

Use the arrows to step through the build. Same content as the live walkthrough, captured for after-the-session reference.

← →

Paste this expression. Or import the JSON.

The AI Note field on the Compose Placeholder Set node uses this expression. Each row gets a different note because {{ $json['Job Title'] }} and {{ $json.Company }} resolve to the current row.

Pending AI review for {{ $json['Job Title'] }} at {{ $json.Company }}.

Same shape as CP1. Wait + LLM in the loop body.

We add three nodes inside the loop. The Wait node throttles the loop so 300 of us together don't hammer the OpenAI proxy. The Basic LLM Chain calls the model with our recruiter prompt. An OpenAI Chat Model sub-node hangs off it with the Hackathon 2026 credential.

CP2 build, screen by screen.

Adding the LLM, the model sub-node, the credential, and the prompt. Use the arrows to step through.

← →

Paste the recruiter prompt. Or import the JSON.

Drop this into the Prompt field of the Basic LLM Chain node, with Source for Prompt set to Define below. The LLM gets the full row context via the $('Loop Over Jobs').item.json references.

You are a senior recruiter reviewing a job posting before it goes live on Indeed.
Write a single concise review note (4-7 sentences, plain prose, no bullets, no JSON, no markdown headers)
covering exactly these points in this order:

1. Salary realism. Is the pay range realistic for this role and location? If not, say why and suggest a realistic range.
2. Inclusive language. Flag any discriminatory or non-inclusive wording. Quote the phrase if found. If clean, say "Language looks inclusive."
3. Description quality. Suggest one specific improvement.
4. End with: "Top qualifications: X, Y, Z."

Be direct and constructive. Do not hedge.

Job Title: {{ $('Loop Over Jobs').item.json['Job Title'] }}
Company: {{ $('Loop Over Jobs').item.json.Company }}
Location: {{ $('Loop Over Jobs').item.json.Location }}
Pay Range: {{ $('Loop Over Jobs').item.json['Pay Range'] }}
Description: {{ $('Loop Over Jobs').item.json['Job Description'] }}

Structured Output Parser is "ask nicely". Native is constrained.

n8n's Structured Output Parser injects formatting instructions into your prompt and parses the model's text. Auto-fix silently doubles your LLM calls when it fails. Indeed-recommended pattern: HTTP Request against the LLM proxy with a JSON schema, so the model literally cannot produce invalid output.

Parser node: ~90-95% schema compliance, 200-500 prompt tokens of formatting instructions, auto-fix doubles cost on failure with no metric to tell you.

Native via HTTP Request: 100% schema compliance, structurally guaranteed at token generation, zero prompt overhead. Works for OpenAI and Gemini through the Indeed LLM proxy.

Indeed wiki: Extracting structured data from LLM responses in n8n has copy-paste-ready HTTP Request examples for OpenAI, Gemini, and Anthropic.

Two new nodes, on the loop's other branch.

The loop has two outputs. So far we have only used the loop branch. Now we wire the done branch (which fires once, after the last row) to a tiny chain: Set builds the summary text, Slack DMs it to you. One DM per run, not 300.

Find your Slack Member ID, then paste it in.

1. Click your avatar (lower-left in Slack) → Profile.
2. Click the three dots next to Edit status.
3. Choose Copy member ID. It looks like U03RU3PLYHK.
4. In the Slack node: Send Message To → User, User → By ID, paste the ID. Set Message Type to Simple Text Message.

App will DM you from: 2026 Hackathon Training. The Hackathon 2026 Slack credential is pre-configured for the workshop, participants don't have to create their own app.

Why By ID, not the dropdown? Indeed's Slack workspace has thousands of users; the dropdown gets throttled and never loads. By ID is the reliable path.

← →

CP3 build, screen by screen.

Wiring the done branch, composing the summary, configuring the Slack node. Use the arrows to step through.

← →

Paste the summary message. Or import the JSON.

On the Compose Summary Set node, add a string field called summary_message with this expression. The Slack node's Message Text then references {{ $json.summary_message }}.

{{ $('Limit to First 3 (Demo Cap)').all().length }}

:white_check_mark: *Job Review Complete*

Reviewed *{{ $('Limit to First 3 (Demo Cap)').all().length }}* job postings. The AI notes have been written back to the sheet.

:link: <https://docs.google.com/spreadsheets/d/1onDtz4TeM0SmJD7-QCV_QzCJXyQptbt4YnAAbwJaew4/edit?gid=921304816|Open the reviewed jobs sheet>

Six places this workflow can go next.

You built the spine in an hour. The hackathon is for putting muscle on it. Each of these is a half-day project, pick one, fork your CP3 workflow, ship it.

Reach for Code last, not first.

Most workflows that reach for the Code node could have used Set, IF, or a native integration instead. The test is simple: "is there a built-in node that already does this?" Usually, yes.

Hackathon ≠ production. Three reminders.

Today's workflow is a learning demo. If you take it any further, three guardrails apply.