Build your first AI role-fit assistant

Mission

Help a data or AI candidate build a portfolio artifact that proves they can turn an ambiguous AI idea into a useful, bounded product workflow.

Product boundary

This lab teaches a personal portfolio assistant for analyzing a single job description. It does not teach job aggregation, ranking, marketplace matching, recruiter workflows, or proprietary Dataaxy product logic.

Before you start

Access to any chat-based LLM; a free tier is enough and no API key is required
One job posting you genuinely care about for the independent exercise
Three to five lines of candidate evidence from projects, tools, or work experience
No coding is required in modules 1 and 2; implementation starts later in the course

Objective

Define a small role-fit assistant without building a job board, marketplace, scraper, or Dataaxy-style matching product.

Primary source

Use this guide while comparing the vague prompt with the explicit task, evidence rules, and response contract in the worked example.

OpenAI Prompting Guide

Start with a job-to-be-done

A strong AI engineering portfolio starts with a concrete workflow. In this course, the workflow is not "chat with my resume" and it is not "rebuild a job board." It is helping a candidate understand one role well enough to decide what to improve next.

The assistant should behave like a focused analyst. It reads a job description, extracts the work that matters, compares it to candidate evidence, and returns practical next steps.

That boundary matters. Dataaxy can later productize matching, job discovery, and recruiter workflows. The course should teach the transferable AI engineering skill without exposing or diluting the platform strategy.

User: a data or AI candidate reviewing job postings
Input: one job description and a lightweight candidate profile
Output: role fit, missing evidence, and a next action
Boundary: no scraping, no ranking marketplace, no recruiter-side automation

Design the first response contract

Before choosing a model or framework, define the shape of a useful answer. This gives you something testable, makes the assistant easier to evaluate, and keeps the project readable for recruiters.

For the first version, use a compact response with five parts: role snapshot, evidence from the job text, candidate strengths, missing proof, and a recommended next action.

This response contract turns the project from a demo into a product artifact. A recruiter can inspect whether the candidate understands evidence, constraints, and user value.

Keep every section short enough to scan
Use evidence from the job text instead of generic advice
Prefer practical next actions over motivational copy
Separate facts from recommendations

Make the prompt inspectable

A portfolio reviewer should be able to read your prompt and understand the product decision behind it. Avoid a prompt that simply asks the model to be helpful. Show the task, the evidence rules, and the response contract.

This is the first habit of applied AI engineering: design the behavior you want, then make the model operate inside that design.

Task: compare a job posting with a candidate profile
Evidence rule: quote the job text for every important claim
Boundary: say "not enough evidence" when the job text is vague
Output: use the same sections every time

Common mistakes

The most common mistake is making the assistant too broad. If it claims to search the market, rank every job, rewrite the resume, and decide whether to apply, it becomes impossible to evaluate and too close to a platform feature.

A second mistake is hiding the product decision inside a long prompt. A good portfolio project makes the trade-offs visible: what the assistant does, what it refuses to do, and why the workflow still creates value.

Too broad: "Find me the best AI jobs"
Better: "Analyze this one role against this evidence"
Too vague: "Give career advice"
Better: "Quote evidence, identify missing proof, suggest one next action"
Do not add a fit percentage: an unsupported score hides uncertainty
Do not rewrite the resume: the artifact should expose reasoning, not keyword stuffing

Worked example

From a vague fit question to an inspectable assistant

Nordwind Mobility and Amina are fictional. The posting is intentionally incomplete so the assistant must distinguish evidence from assumptions.

Sample job posting (fictional)

Analytics Engineer — Nordwind Mobility
Berlin, hybrid-friendly

Nordwind Mobility helps cities understand how people move. We are looking for an Analytics Engineer to build trusted models for product and operations teams.

You will model mobility and revenue data in BigQuery, maintain dbt projects, document core metrics, and partner with analysts and engineers. You will also improve data quality checks, contribute to orchestration workflows, and support ML feature pipelines when needed.

Advanced SQL and hands-on dbt experience are required. Experience with Airflow or a similar orchestration tool is valuable. Python is a plus. You bring a senior mindset and can explain trade-offs to non-technical partners. German is nice to have.

We are a hybrid-friendly team based in Berlin.

Candidate evidence (fictional)

Amina K. — Data Analyst, 3 years
- Uses SQL in BigQuery every day
- Builds Looker dashboards for product teams
- Maintains a personal dbt project in a public repository
- Uses pandas for small Python scripts
- Has not used an orchestration tool in production

Filled user story

As a data analyst considering analytics-engineer roles, Amina wants to understand which missing evidence could block this specific application so she can decide what to build next.

Prompt v1: too vague

This version invites generic advice, unsupported scores, and invented requirements.

Am I a good fit for this job? Give me a score and improve my resume.

Prompt v2: bounded and evidence-led

You are reviewing one job posting for one candidate.

Use only the JOB POSTING and CANDIDATE EVIDENCE provided below.

Return exactly five sections:
1. Role snapshot
2. Evidence from the posting
3. Candidate strengths
4. Missing or uncertain evidence
5. One next action

Rules:
- Quote the job posting verbatim for every requirement you discuss.
- Separate explicit facts from interpretations.
- Write "not enough evidence" when the posting or candidate evidence is unclear.
- Do not produce a fit percentage, rewrite the resume, infer salary, or compare other jobs.
- Recommend one concrete next action only.

JOB POSTING:
[paste one posting]

CANDIDATE EVIDENCE:
[paste three to five evidence lines]

Expected assistant response

## 1. Role snapshot
This role centers on analytics engineering in BigQuery and dbt, with collaboration across product, analytics, and engineering.

## 2. Evidence from the posting
- Core requirement: "Advanced SQL and hands-on dbt experience are required."
- Orchestration exposure: "Experience with Airflow or a similar orchestration tool is valuable."
- Python is secondary: "Python is a plus."

## 3. Candidate strengths
- Amina's daily BigQuery SQL experience aligns with "model mobility and revenue data in BigQuery" and the required advanced SQL, although her evidence does not establish the exact depth expected.
- Her public dbt project aligns with "maintain dbt projects," although the posting does not state the expected project scale.
- Her product-team work is relevant to "partner with analysts and engineers," but her evidence does not show engineering collaboration directly.

## 4. Missing or uncertain evidence
- Nice-to-have orchestration evidence is missing. Amina has no production evidence for Airflow or an alternative, but the posting calls this experience "valuable," not required.
- Seniority: not enough evidence. The posting says "senior mindset" but gives no years or ownership threshold.
- Salary: not enough evidence. The posting gives no compensation range.

## 5. One next action
Add an Airflow or Dagster workflow around the existing dbt project and document one retry, one failure alert, and the trade-off behind the schedule.

Assistant brief checklist

Name the user and the moment where the assistant helps
State the input the assistant receives
Define the answer format before writing prompts
Include evidence rules so the answer is grounded in the job text
Describe one failure mode the assistant must avoid
State what the assistant intentionally does not do

Retrieval practice

Answer before you continue

Who is the user, and what decision are they trying to make?
What evidence should the assistant quote from the job description?
What answer shape would make the output easy to compare across jobs?
Which part of this idea would become a platform feature if Dataaxy built it internally?

Exercise

Write a one-page assistant brief, draft three prompts, and score the assistant output against a concrete review rubric.

Build the first assistant brief

Use one public job description and one lightweight candidate profile. Draft the assistant brief before writing any code.

Write a one-sentence user story for the candidate workflow
List the exact inputs the assistant can use and the inputs it must ignore
Define a five-part answer format: role snapshot, evidence, strengths, gaps, next action
Write one prompt that forces the assistant to cite evidence from the job text
Add a "not enough evidence" rule for vague job descriptions
Write a short product boundary note explaining why this is not a job board

Reveal the feedback checklist

The user story names a concrete candidate decision
The input fields are observable and not invented by the model
The answer format can be reused on a second job posting without redesign
The prompt asks for evidence instead of generic career advice
The boundary note keeps Dataaxy platform logic out of the portfolio project

Reveal the expected artifact

A recruiter-readable assistant brief

Publish the brief, final prompt, and one reviewed answer in a repository or public gist. A reviewer should understand the product decisions without running the assistant.

One-page brief template

# AI Role Fit Assistant

## User and decision
As a [candidate type], I want to understand [one decision] so I can [one action].

## Allowed inputs
- One manually provided job posting
- Three to five candidate evidence statements

## Inputs intentionally ignored
- Other job postings
- Market salary estimates
- Hidden candidate or recruiter data

## Response contract
1. Role snapshot
2. Evidence from the posting
3. Candidate strengths
4. Missing or uncertain evidence
5. One next action

## Evidence and refusal rules
- Quote the posting for every discussed requirement
- Say "not enough evidence" instead of guessing
- No fit score, resume rewrite, job ranking, or salary inference

## Known failure mode
The assistant may treat a nice-to-have as mandatory. Review every classification against the original quote.

Submission checklist

Include:
- The final prompt with the five-section contract
- One sample input using fictional or permitted data
- One complete sample answer
- At least one explicit uncertainty or refusal
- A short note explaining what the assistant does not do

Review rubric

Check whether the brief meets the bar

Check whether the brief meets the bar
Criterion	Strong	Weak
Concrete user decision	The user story names one candidate, one role, and one decision.	The goal is generic career advice or job discovery.
Evidence rule	The prompt requires a verbatim quote for every discussed requirement.	The prompt only asks the model to be accurate or helpful.
Reusable contract	The same five sections work on a second posting without edits.	The sections contain Nordwind-specific fields or wording.
Uncertainty	The answer says "not enough evidence" and refuses unsupported claims.	The answer fills gaps with plausible assumptions.
Product boundary	The brief names concrete refusals: no ranking, scraping, salary inference, or resume rewrite.	The boundary is a generic disclaimer with no observable refusal.
Next action	One action addresses the most important missing proof and produces evidence.	The response ends with several generic learning suggestions.

Artifact to ship

Assistant brief, final prompt, and one sample answer

Ship it in GitHub

Continue in the GitHub starter

Use the public Dataaxy template to keep your prompt, schema, examples, rubric, and project checklist in one recruiter-readable repository.

Open the portfolio starter kit

Ask for feedback

If you are working with an AI coach or mentor, ask it to review your artifact against the checklist before moving to the next module.