34 Applicants, 2m 40s, $0.11: The Telegram Hiring Bot

By Lazar Milicevic · Published July 22, 2026 · 8 min read

Forty responses sitting in a Google Sheet you haven't opened. Every row is a human, and you know it. That guilt is what pushes solopreneurs into $200/month ATS subscriptions they don't need — when what they actually need is a bot that reads the sheet, ranks it, and pings their phone with the top three.

Here's the exact stack I ran last month for a client hiring one contractor: 34 applicants scored in 2 minutes 40 seconds, total Claude spend $0.11, zero recurring infrastructure cost. No Greenhouse, no Workable, no monday.com. Let me walk you through the build and — more importantly — where it breaks.

Why an ATS is the wrong tool for hiring one person

Greenhouse, Workable, Lever and monday.com's recruiting product are genuinely well-built software. They are also built for a hiring manager filling twelve roles this quarter with three recruiters supporting them. Pipeline stages, scorecard templates, interview-kit assignments, EEOC reporting — none of that maps to a solopreneur hiring their first VA.

The job to be done at your scale is smaller and sharper:

• Read every response so nobody gets ghosted.
• Rank them against a rubric you actually care about.
• Surface the top three to your phone so you can act during lunch.

That's a scoring problem plus a notification problem. It is not a pipeline problem. Paying $149-$399/month for pipeline software to solve a scoring problem is like buying Salesforce to track your gym buddies.

The stack, deliberately boring

Four pieces, all of which you probably already have accounts for:

• Google Form for intake. Free, mobile-friendly, everyone knows how to fill one out.
• n8n (self-hosted or cloud) as the orchestrator.
• Claude Haiku via API as the scoring model.
• Telegram bot as the notification layer.

No database. No CRM. No vector store. The Google Sheet backing the form is the database. If you self-host n8n on a $6 VPS you already run, marginal infrastructure cost is $0. If you use n8n Cloud, it's $20/month and covers 50 other workflows too.

The Claude API call is the only per-run cost. At Haiku pricing, 34 applicants with ~800 tokens in and ~60 tokens out per applicant lands at roughly $0.11 for the batch. I've re-run this pattern for four separate clients — the cost stays in the $0.08–$0.15 range for batches under 50.

Step 1: The form and the sheet

Keep the form short. Every additional field cuts your response rate and gives the model more noise to weigh.

Fields I use:

• Name
• Email
• Timezone (dropdown, not free text)
• One link — portfolio, GitHub, or LinkedIn
• One open-ended question: "Describe the last project you shipped end to end. Two paragraphs max."

That last question is the whole game. The score is 80% driven by what they write there. Vague answers ("I've done many projects across various industries") get flagged. Specific answers ("Migrated a Shopify Plus store from Liquid to Hydrogen, cut LCP from 4.1s to 1.6s, went live March 14") score high.

Link the form to a Google Sheet in the form's Responses tab. That sheet becomes your audit trail — searchable, exportable, and legally defensible if anyone ever asks how you evaluated candidates.

Step 2: The n8n workflow

Two workflows, not one. Separating them keeps debugging sane.

Workflow A — Score on new response:

1. Google Sheets Trigger → On row added, pointed at the responses sheet.
2. HTTP Request → Claude API (scoring prompt below).
3. Code node → parse the JSON response.
4. Google Sheets → write score and reasoning back to the row.

Workflow B — Shortlist push (runs every 2 hours, or manually):

1. Google Sheets → read all rows where score >= 70 AND notified != true.
2. Sort by score descending, take top 3.
3. Telegram → send formatted message with inline buttons.
4. Google Sheets → mark those rows notified = true.

I use the Sheets trigger instead of the native Google Forms trigger because Sheets lets you reprocess a batch by clearing the score column. Forms trigger fires once per submission and gives you no replay path when you tweak the rubric.

Step 3: The scoring prompt (this is where it lives or dies)

Four weighted criteria. The weights are the opinionated part — change them for your role.

You are evaluating a job application for a [ROLE] contractor role.
Working hours: 9am–6pm US Eastern.

Score the applicant 0–100 using these weighted criteria:

1. Relevant experience (weight 40)
   - Concrete, verifiable projects in the required stack
   - Deduct heavily for vague or generic claims

2. Portfolio link quality (weight 25)
   - Does the link work, is the work real, is it recent (<18 months)

3. Timezone overlap (weight 20)
   - 6+ hours overlap = full marks
   - 3–5 hours = half marks
   - <3 hours = zero

4. Red flags (weight 15, subtract)
   - Obvious AI-generated fluff ("leverage synergies", "passionate about")
   - Mismatched claims (senior title, junior detail)
   - No specifics, only adjectives

Return JSON only, no prose:
{
  "score": <int 0-100>,
  "reasoning": "<one line, max 20 words>"
}

Application:
Name: {{name}}
Timezone: {{timezone}}
Link: {{link}}
Project: {{project_answer}}

Two constraints matter here. First, JSON only, no prose — otherwise n8n's Code node has to strip markdown fences and you'll spend an afternoon debugging edge cases. Second, one-line reasoning, max 20 words — because the output lands on a phone screen and anything longer gets truncated.

Haiku is the correct model choice. It's fast (sub-second per applicant at this prompt size), it's cheap, and for structured scoring against an explicit rubric it's more accurate than you need. I've tested the same prompt on Sonnet — scores shifted by an average of 3.2 points, ranking order changed on 2 of 34 applicants. Not worth 12x the cost for a shortlist you're going to eyeball anyway.

Step 4: The Telegram message that actually gets read

The BotFather setup is boring — /newbot, name it, grab the token, paste into n8n's Telegram credentials. What matters is the message format.

Bad format (what most tutorials show):

New candidates:
1. Sarah Chen - 88/100 - Sarah is an experienced developer who has worked on many projects including a recent Shopify migration and appears to have strong skills in the required areas based on her portfolio which demonstrates good work.
2. Marcus Ali - 82/100 - Marcus has...

That's unreadable on a phone. Here's the format I ship:

🥇 Sarah Chen — 88
Shipped Shopify Plus → Hydrogen migration solo, LCP 4.1s→1.6s
sarahchen.dev
[Reply] [Reject]

🥈 Marcus Ali — 82
Built 3 internal tools in n8n for a 40-person agency, screenshots verified
github.com/marcusa
[Reply] [Reject]

🥉 Priya Shah — 76
6 months freelance React, one live case study with metrics
priya.build
[Reply] [Reject]

Rank, name, score, one-line reasoning, link, two buttons. That's it. The Reply button hits a second n8n webhook that drafts a Gmail from a template with the candidate's name and email pre-filled. The Reject button marks the row status = rejected in the sheet. Both buttons take one tap.

Step 5: The real run, unedited

Client posted a contractor role in a Slack community on a Tuesday. Form went live at 9am. By Wednesday noon, 34 responses.

• Workflow A had been firing on every submission for 27 hours, so all rows already had scores.
• At 12:04pm we ran Workflow B manually.
• 12:06:40pm — one Telegram message, three names, three scores, three links.
• Client tapped Reply on ranks 1 and 2 during lunch. Reject on 8 low-scorers he skimmed later that evening.
• Both replied-to candidates responded within 4 hours. He hired rank 1 by Friday.

Setup time when we built it together over a screenshare: 45 minutes, most of which was BotFather auth and getting the Claude API key into n8n credentials.

Where this breaks (the honest section)

Every tutorial that skips this is lying to you.

• PDF resumes. If your form asks for a PDF upload, Haiku will not parse it reliably from the base64 blob. Either strip the file field and require a link instead, or add a Google Drive → PDF-to-text step before scoring. I chose the first option — links filter for people who have a portfolio worth showing.
• Non-English applications. The rubric works fine for English, French, Spanish. For CJK languages the "red flags" criterion is unreliable because "AI-generated fluff" patterns don't transfer. Run those through Sonnet instead or add a language-specific rubric.
• Bias. The model will happily score based on signals correlated with things you shouldn't hire on. I explicitly remove name and photo from the prompt payload and only send timezone, link, and project answer. Even then, writing style correlates with education and native language. This is a shortlist tool, not a decision tool — you still read the top 3 yourself.
• Volume ceiling. Above ~200 applicants per batch, the "top 3 to Telegram" pattern breaks because rank 4–10 are often within 5 points of rank 1–3. At that scale switch to a paginated Telegram flow or a simple web view.
• Legal. In the US, EU, and UK, automated candidate screening has disclosure requirements in some jurisdictions (NYC Local Law 144, EU AI Act's high-risk classification for employment). If you're hiring at any scale, check the current rules for your state or country and add a disclosure to the form. This post is not legal advice.

Why bizflowai.io helps with this

The Telegram hiring bot is one of a dozen "solopreneur-scale" workflows I build for clients at bizflowai.io — small, opinionated automations that replace a $200/month SaaS with a $0.11/batch n8n workflow. Hiring is one; inbound lead triage, invoice reminders, competitor price monitoring and support-ticket routing follow the same pattern. The point isn't the bot, it's the mindset: match the tool to your actual volume, not to the volume the SaaS vendor built for.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.