Flashcards

Create one /workspace/<slug>.json file. JSON is the canonical deck; do not write a separate Markdown summary.

Contract

Use exactly this closed version-one shape:

{
  "schema_version": 1,
  "title": "Deck title",
  "cards": [
    {
      "front_text": "One focused recall prompt",
      "back_text": "A sufficient, concise answer",
      "hint_text": "Optional hint with inline math: \\(x^2\\)"
    }
  ]
}

• Include 15–100 cards. Use a requested count within that range; otherwise choose 30–50 based on the source depth. Use mixed difficulty when unspecified. Do not ask for confirmation.
• Test one distinct recall target per card. Plan and deduplicate the targets before writing JSON; do not repeat, paraphrase, reverse, or split the same fact merely to reach the minimum.
• Report insufficient material only after finding fewer than 15 distinct recall targets in the available source and topic. Do not invent filler.
• Title: 1–200 characters. Front: 1–4,000. Back: 1–12,000. Hint: 1–2,000 when present.
• Keep each front atomic and each back sufficient without testing multiple unrelated facts.
• Card fields are plain text. The only formatting syntax allowed is LaTeX: use \(...\) for inline math or \[...\] for display math. Escape each backslash as \\ in JSON. Keep delimiters and braces balanced; do not nest math delimiters.
• Omit hint_text when no useful hint exists; never use null.
• Do not add fields, change schema_version, duplicate fronts, or use raw HTML or Markdown in card content.
• Write UTF-8 JSON without comments or a byte-order mark.

Verify and save

Call:

verify_artifact(path="/workspace/<slug>.json", format="flashcards")

Fix all blocking findings together, rewrite the JSON once, and reverify the exact path. If a blocker remains, stop and explain it instead of looping.

After verification returns status="verified", call:

save_artifact(path="/workspace/<slug>.json", title="...")

Do not pass markdown_representation; the backend derives trusted Markdown from the exact receipt-bound JSON.

For revisions, edit the restored JSON from primary_path, write the complete deck to expected_output_path, verify that path as flashcards, and save with the returned artifact_id and expected_generation. Do not edit the derived markdown_path or reconstruct the deck with vision.