Guide

During the hackathon you work in a small team (1–3 people, formed freely on Day 1) on one simulation method. You work with two things: an AI coding agent — a chat assistant that runs in your terminal and can read files, write code, and run programs — and the harness, the shared GitHub repository QuantumBFS/quantum.harness, where an expert for each method has collected the key papers, worked examples, and step-by-step instructions that teach the agent to run that method correctly.

Your goal for the week: reproduce a published result for your method, then go beyond it with a challenge of your own, and submit everything as a pull request — GitHub’s mechanism for proposing your changes back to a shared repository, and the form in which advisors review your work.

This page walks through the week in order. Each ▶ Prompt box is text to paste into your agent (Claude Code, Codex CLI, or OpenCode — see Setup); hover over a box for a copy button.

// Deadlines

// Methods

Each card names the method, the expert who curated that track, and the command that loads the method’s instructions in your agent.

Track folder names in the repository (also used in your submission path and PR title): ed, mps, peps, qmc, qcs, polyopt, agent-kb.

Source of truth: the QuantumBFS/quantum.harness method table.

// Setup — before you arrive

Install one AI coding agent and log in before July 26 — there is no installation session on Day 1. You also need a free GitHub account.

To verify it works, start the tool in any folder and ask it to read a file.

Next, install two helpers. superpowers is a collection of skills — reusable instruction files that teach the agent careful working habits, such as planning before coding and testing before claiming success. gh is GitHub’s command-line tool; the agent uses it to copy the repository and open your pull request.

Install https://github.com/obra/superpowers and gh (GitHub CLI) if not already installed.
Authenticate me to GitHub via gh auth login if I'm not logged in.

If your agent lets you pick a model with /model, choose the strongest one — Claude Code: Opus 4.7 (high effort); Codex CLI: GPT-5.4 (xhigh effort); OpenCode: route to one of the above via your configured provider.

// Get the harness (Day 1)

You never edit the shared repository directly. Instead you fork it — make your team’s own copy on GitHub — and work on a branch, a named line of work inside that copy. The prompt below also runs make skills, which installs the harness’s skills so that commands like /beginner-training become available.

Fork https://github.com/QuantumBFS/quantum.harness to my GitHub account using gh repo fork.
Clone the fork to ~/code/quantum.harness and add the upstream remote.
Run make skills to install the harness skills.
Create a working branch named group-<team-name> off main.
Verify with git status, git remote -v, gh auth status, and the current branch name.

Pick <team-name> yourself: short, lowercase, hyphens instead of spaces (e.g. wolf-pack). Teams are self-organized, size 1–3, one pull request per team.

// Onboard — reproduce a published result

Each track names a reference paper and the specific figures or tables to reproduce. Matching the published numbers proves that your whole setup — agent, code, environment — works end to end, and earns you the right to move on. Start with:

/beginner-training

Run /beginner-training. It offers five guided tracks — setup check, reproduce a paper, literature survey, develop code like an expert, and go beyond — and walks each one step at a time, explaining before it runs anything. No prior git, GitHub, or Julia experience is assumed; that is the training’s job. For the reproduction, take Track 2: it lists every method track’s reference paper, lets you pick one, and then guides the reproduction step by step. When it finishes, note what matched the paper, what didn’t, and which hardware and software you ran on — you will need all of this for your report.

// Take on a challenge

A challenge is a well-posed question that goes beyond the reference paper: a larger system, a different observable, a sharper bound, a better algorithm. Pick an accepted idea from the Challenges page, or run /challenge to survey the literature and shape one of your own. Either way, confirm your plan at the on-site help desk before spending days on it.

To think through a listed challenge with your agent:

Read challenge issue #<N> from https://github.com/QuantumBFS/quantum.harness/issues/<N>.
Brainstorm with me: surface 3–5 distinct angles, name the riskiest assumption in each, and recommend one to start with.

// Submit — pull request by Day 4 morning

Your committed work lives in your fork under tracks/<track>/solutions/ — the same place the harness skills put it. Generated data and figures land in tracks/<track>/results/, which stays out of git; reviewers re-run your result from your reproduction prompt instead. Your pull request must contain three things:

1. An improvement to the harness itself — for example, sharpen a method skill or fill a gap in the knowledge base that you hit during the week.
2. Your challenge solution under tracks/<track>/solutions/ — the scripts that produce your result, plus a README that pitches what you did and why it matters.
3. One reproduction prompt, posted as a comment on the pull request and generated with the harness’s challenge-report skill, so a reviewer can re-run your result from scratch.

Create a pull request to QuantumBFS/quantum.harness:main following the submission guideline at https://giggleliu.github.io/summer-school-2026/challenge-ideas#submission-guideline.

After the pull request opens, check its page on GitHub: the automated checks (labelled “CI”) must all be green — passing — by the Day 4 morning deadline.

// Day 5 — advisor review & showcase

Use Day 4 afternoon to make your pull request easy to review: rerun your demo from a clean start and reread your README with fresh eyes. On Day 5 morning each track meets in its own room, and the advisor decides at their own discretion which results to back with their sponsor-provided Mac mini award(s).

// Getting help

• Zulip — harnessing-quantum.zulipchat.com: announcements, help-desk Q&A, and per-track threads. Your invite arrives with the acceptance email.
• Help desk (on site) — consultation, schedule guidance, and ad-hoc support, staffed by teachers and TAs.
• Expert office hours — scheduled during the week; see the Program.

// Learn more

Learning:

• The Complete Guide to Building Skills for Claude (PDF)
• Agentic Coding — MIT Open Course (Video)

Optional tools & skills: the Resources page lists MCP servers, command-line tools, and skill collections worth knowing.