The setting

Quillstream sells a developer documentation platform and is building Quill Assistant, a retrieval-grounded support copilot that answers customer questions from the product docs. You are an AI engineer on the assistant team.

Every mission on this path happens at the same company, so context carries over the way it does in a real job: the data you cleaned in mission two is the data the finance lead questions in mission four.

The missions

1. prompt designstarter

Rewrite the Quill Assistant system prompt so it stops guessing

Quill Assistant is confidently wrong too often. Product manager Dana Osei traced most complaints back to the system prompt, which tells the model to be helpful and confident and says nothing about staying inside the docs.

You deliver: A revised Quill Assistant system prompt plus a rationale mapping each change to a current weakness.

Scored on: Grounding rule, Citation and refusal, Removes the confident-guessing instruction, Rationale is mapped.

Working from: assistant_context.md, current_prompt.md.

2. RAG groundingstarter

Catch the ungrounded claims in a Quill Assistant answer

A grounding check surfaced one assistant answer that reads well but says more than the retrieved docs support. Before it becomes a template for other answers, you need to pull it apart.

You deliver: A grounding review flagging unsupported claims plus a rewritten, cited answer.

Scored on: Flags the ungrounded claims, Confirms grounded parts, Correct rewrite, Cites sources.

Working from: retrieval_trace.md.

3. tool usecore

Fix the two tool-use mistakes in a Quill Assistant run

Quill Assistant can call four tools, but a review of one conversation shows it using them poorly: answering product questions from memory and opening tickets it should not. Engineer Lars needs a decision policy the model can follow and the two specific fixes.

You deliver: A tool-use design note with a decision policy and corrected sequences for the two logged mistakes.

Scored on: Clear decision policy, Catches the memory answer, Catches the premature ticket, Corrected sequences.

Working from: tool_catalog.md, agent_run_log.md.

4. evaluationcore

Read the Quill Assistant eval and find the failing cluster

The latest evaluation run scored 9 of 14 test cases as passes. A 64 percent pass rate is a number, not an insight. Before the team celebrates or panics, someone has to find where the failures actually are.

You deliver: An eval report with the pass rate, the failing cluster, a root-cause hypothesis, and ranked fixes.

Scored on: Correct pass rate, Finds the failing cluster, Root-cause hypothesis, Ranked fixes.

Working from: eval_results.csv.

5. governance checksstretch

Map the Quill Assistant red-team findings to policy

Before Quill Assistant ships to all customers it needs a governance sign-off. The red-team run produced three incidents, and each one breaks a rule in the governance policy. Leadership wants to know exactly what broke and what control stops it.

You deliver: A governance review mapping each red-team incident to policy with a control and a go or no-go call.

Scored on: Covers all three incidents, Correct policy mapping, Concrete controls, Go or no-go call.

Working from: governance_policy.md, redteam_log.md.

How the scoring works

Each deliverable is graded against the rubric written for that mission. Separately, every mission on every path is graded on how you used AI, against the same four criteria:

  • Understood the task. The learner framed the goal for the assistant clearly instead of pasting the brief and hoping.
  • Grounded in the material. The learner directed the assistant into the provided files and based the work on them, not on invented facts.
  • Verified the output. The learner checked claims, numbers, or coverage against the source material before submitting.
  • Iterated with judgment. The learner refined weak parts of the draft with specific follow-ups rather than accepting the first answer.

Both scores, with the work behind them, go on your proof profile. That is what makes a claim like "I can use AI for ai engineering" something an employer can check.