| name | beginner-training |
| description | Use when a new harness user (assume undergraduate level) starts the beginner training and needs to choose which track to run — asks which of the five tracks (setup check, reproduce a paper, literature survey, develop code like an expert, go beyond) to take, then walks that track as a guided, confirm-gated tutorial that explains every step and waits for the student to confirm understanding before running anything. Triggers on "start the training", "beginner training", "/beginner-training", "which track should I take", "I'm new here, teach me the harness". |
beginner-training
Overview
The launcher for the harness beginner training. It does two jobs: (1) ask
which track to take and route there, and (2) carry the Teaching Protocol
below — the rules for how every track is run. The track pages under
skills/beginner-training/tracks/ hold each track's step content; this skill
holds the selector and the protocol.
Assume the student is at undergraduate level. Do not assume they have used
git, GitHub (gh), Julia, a literature-survey tool, or quantum many-body
jargon before. Define terms the first time they come up (what is a symmetry
sector? DMRG? a residual? a PR?), and explain why a step matters, not just
how to run it.
Teaching Protocol — follow this for EVERY step of EVERY track
This is a training, not a script. Teach one step at a time, and never run
a command before the student confirms they understand it. For each step:
- Announce where we are: "Step N of M — ."
- Explain what will happen, in plain language a first-year student follows.
- State the purpose — what they're learning, or why researchers work this way.
- Show the exact command and what a correct result looks like (the
observable expected output).
- Stop and wait. Invite them to confirm or ask: "Reply
ready and I'll
run it — or ask me anything first." Do not run anything until they
confirm. If they ask a question, answer it, then re-offer.
- Run it (only after they confirm), show the real output, and explain
what actually happened, tying it back to the purpose. If it failed,
diagnose it with them before moving on.
- Move to the next step.
Hard rules
- One step, one confirmation. Never batch-run steps.
- Never skip the explanation to save time — the explanation is the training.
- Define jargon on first use (what is a Hamiltonian sector? a bond dimension?
a pull request?).
- If the student says "just do it" / "skip ahead," you may proceed, but still
say what each step did as you go.
- Every track ends with a Checkpoint, run this same confirm-gated way, whose job
is to prove trustworthiness — via a real negative control (a deliberately
bad input that must fail) where that fits, or a cheaper integrity check
(self-contained, placeholder-free, count matches what actually ran) where a
staged negative control would be artificial. Run whichever the track specifies
("now let's prove the check actually has teeth"). Do not demand a literal
negative control on a step whose real safeguard is an integrity check.
Step 1 — Ask which track to take (do this first)
Present the track selector. The recommended order is 1 → 2 → 3 → 4 → 5, but any
track can be taken first and each stands alone. AskUserQuestion allows four
options, so offer: Recommended full path (1→2→3→4→5), Track 1,
Track 2, and a catch-all Another track (3, 4, or 5). If the catch-all
is picked, ask one follow-up question offering tracks 3, 4, and 5. The tracks:
- Track 1 — Setup check →
tracks/track1-setup.md. No prerequisite.
Recommended before any other track.
- Track 2 — Reproduce a paper →
tracks/track2-reproduce.md. Precheck:
ls tracks/*/README.md lists the method tracks, and pdf-render is available
(.venv/bin/python -c "import pymupdf4llm"). Scans the method-track READMEs,
lets the student pick one, then hands off to /reproduce-paper at beginner
tier with teaching mode kept on.
- Track 3 — Conduct a literature survey →
tracks/track3-survey.md.
Precheck: .venv/bin/python -c "import pymupdf4llm" succeeds. Builds a small
reference library in .knowledge/ with /download-ref and /survey.
- Track 4 — Develop code like an expert →
tracks/track4-develop.md.
Precheck: gh auth status succeeds and gh repo view QuantumBFS/qsym-rs
succeeds, plus the superpowers loop resolves
(skills/brainstorming/SKILL.md, skills/writing-plans/SKILL.md,
skills/requesting-code-review/SKILL.md all exist). The learner must
explicitly pick which starter issue to work on — the trainer must not open,
assume, or drive any specific issue before the learner picks it.
- Track 5 — Go beyond →
tracks/track5-beyond.md. The capstone. Precheck:
a completed reproduction to extend (Track 2 done, or the student names a
reproduced result). Runs /challenge → /challenge-report.
Plus Recommended full path → run 1 → 2 → 3 → 4 → 5 in order. Any track is
selectable first; do not force the order.
Step 2 — Run the chosen track under the Teaching Protocol
- Open the chosen track page. Give the student the track's Goal and
Before you start overview, and confirm they're ready to begin.
- Run that track's scoped precheck as the first teaching step(s) — explain
what you're checking and why, then wait, then run. If a check fails, walk the
student through the matching Track 1 fix before continuing.
- Walk the track's Steps one at a time, each following the Teaching Protocol.
- Run the Checkpoint the same confirm-gated way — proving trustworthiness via
its negative control or integrity check, whichever that track specifies —
and make sure the student understands why the checkpoint passed.
- When the checkpoint passes, congratulate them and offer the next track in the
recommended order.