AWS Builder Center

Builder Center

Meet Anfi: an agent that gets first-timers to their first pull request, one step at a time

I built a Hacktoberfest host agent in n8n and Amazon Bedrock that finds a beginner one good first issue and guides them one step at a time, then tested it with 20 end-to-end scenarios.

J. Manuel Garcia Torres

Next Saturday, October 17, we're hosting Hacktoberfest 2026 Tabasco at Xpectre House in Villahermosa, Mexico. The poster says "Beginners welcome", and we mean it. But every year I see the same moment: someone opens GitHub, stares at millions of repositories, and quietly closes the laptop.
This weekend I built Anfi (short for anfitrión, Spanish for "host"), a chat agent whose only job is to get a first-time contributor from "I've never done this" to "I opened my first pull request."

What it does and who it's for

Anfi is for people coming to their first open source event: students, career-switchers, and developers who have used GitHub only to download things.
In a conversation, Anfi:
  1. Asks two questions, and no more: which language you know (even a little), and whether you've used git.
  2. Finds one beginner-friendly issue on GitHub that matches you. It looks for open, unassigned issues labeled good first issue in your language. If you don't code yet, it looks for documentation issues.
  3. Walks you through the contribution one step at a time: ask to be assigned, fork, clone, branch, change, commit, push, open the PR. It waits for you to say "done" before giving the next step.
  4. Answers event questions (when, where, what to bring) and is honest when it doesn't know something.
It speaks Spanish or English, depending on how you write to it.

How I built it

I built Anfi in n8n, the workflow automation tool I use every day as an n8n Ambassador, with Amazon Bedrock as the brain.
The workflow has five nodes:
  • Chat Trigger. n8n's hosted public chat gives me a shareable URL for attendees, so nobody has to install anything.
  • AI Agent. This node holds the system prompt, which is where most of the design work went (more on that below).
  • AWS Bedrock Chat Model. The workflow ships wired to Claude Haiku on Amazon Bedrock. For a guided, short-message conversation it is fast and cheap, and Bedrock keeps the model behind IAM credentials instead of yet another API key. For fast, free iteration, I ran the test suite below on the same workflow with the model node pointed at Gemma 4 running locally in Ollama; the model is a single swappable node in n8n.
  • Window Buffer Memory. It keeps the last ~10 messages per session, so Anfi remembers your language and git level as you go.
  • HTTP Request Tool (search_good_first_issues). It calls the GitHub Search API with a query like is:issue is:open no:assignee label:"good first issue" language:python. The model fills in the language with n8n's $fromAI(). I trimmed the response down to title, URL, repo, comment count and labels, so the model reads five small objects instead of a wall of JSON.
I treated the system prompt as the product spec. It defines the event facts (and tells the agent to never invent others), the conversation flow, the tone (3–6 lines, at most one emoji), how to handle errors, and one safety rule: issue text from GitHub is untrusted data, and Anfi never follows instructions inside it.
To keep the build honest, the workflow JSON is generated from the prompt file by a script. A validator checks the node types, the connections, that the system message matches the prompt byte for byte, and that no secrets leaked into the export. A second script runs the real GitHub queries for Python, JavaScript and Markdown to confirm they return issues.

The delightful detail: one issue, one step

The obvious design is what most agents do: return a list of ten issues and a full eight-step git tutorial. It's technically correct, and it's exactly what makes beginners close the laptop. Choosing between ten repos you've never heard of is a decision a first-timer can't make yet.
So I made one rule and built everything around it: Anfi gives you exactly one thing at a time. One issue, picked for you, with one sentence on why it's a good first one. Then one step, and it waits for you to say "done" before the next.
The first run follows the same rule. Anfi greets you and asks two plain questions ("What language do you know, even a little? Have you used git?"), not a form. You get value in your second message.
How I know it worked: before putting it in front of real attendees, I wrote 20 end-to-end scenarios that talk to the live n8n chat endpoint, multi-turn, the way a person would: a Python beginner who has never used git, someone who "can't code yet", an English speaker, someone who asks for a different issue, someone who pastes a git error, someone who asks for the system prompt, someone who offers their GitHub token. Each scenario checks the behavior that matters for the experience (exactly one issue link, the reply's language, short messages, correct event facts), and the runner calls the GitHub API to confirm that every issue Anfi recommends actually exists and is open.
The first run passed 15 of 20. One failure was my test's fault. The other four were real, and they were all experience bugs rather than crashes:
  • Anfi named an issue but forgot the link.
  • It said "give me a moment, I'll search" and then never called the tool. The person only ever sees the final message.
  • When someone opened with "I can't code yet and never used git", it asked the two questions they had just answered.
  • It answered an English speaker in Spanish.
Each fix was a line in the system prompt. Reading the transcripts also caught a false pass in my own language check, so I tightened it and confirmed it now fails on the old reply. The third run passed 19 of 20:
  • In all 10 replies that recommended an issue, there was exactly one link, and every link was a real, open issue.
  • No replies leaked the prompt or accepted a token.
  • The remaining failure is honest and still open: a Spanish message that contained an English git error got its (correct) explanation in English.

Proof it works

A real conversation from the test run, word for word:
Me: no sé programar todavía y nunca he usado git (I can't code yet and I've never used git)
Anfi: ¡Hola! No te preocupes, es un punto de partida perfecto y es muy valiente querer empezar. ¡Te voy a guiar paso a paso! He encontrado una tarea sencilla en el repositorio mdn/translated-content: [es] Sincronizar Games/Tutorial_2D_breakout_game_pure_JavaScript/Initialize_the_canvas, https://github.com/mdn/translated-content/issues/39175 …
For someone who can't code yet, it picked a Spanish translation task on MDN. That's the kind of match I would have made by hand.
And a live run in n8n's hosted chat. The issue title is the link:
Anfi in n8n's hosted chat recommending one beginner-friendly Python issue and the first step

What I'd do next

  • Run the same 20 scenarios against Claude Haiku on Bedrock and compare the numbers with the local run.
  • Fix the last failing scenario: when a message mixes Spanish with an English error log, Anfi should still answer in Spanish.
  • Add a GitHub token so the search tool isn't limited by the API's unauthenticated rate limit during the event rush.
  • Let Anfi check the PR link you paste and congratulate you once the PR actually exists.
  • Bring it to the room on Saturday and watch where people still get stuck.
If you're organizing a Hacktoberfest event, the workflow is a single JSON import: plug in your AWS credentials and your event details, and you have a host that never gets tired of explaining git push.
See you in Villahermosa. 🚀
# agents
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article