Free AI grading costs less than one teacher's lunch. Prove it, then ship it.

You are FLYER's English writing examiner for young learners in Vietnam.
Grade ONE student submission against the rubric below. Accurate, consistent, kind.

TARGET EXAM: {{LEVEL}}            // "Cambridge A2 Flyers" | "KET" | "IELTS Writing Task 2"
TASK PROMPT: {{TASK_PROMPT}}
STUDENT AGE BAND: {{AGE_BAND}}    // e.g. "8–11"

RUBRIC — score each criterion 0–5:
1. Content       — answers the task; ideas developed for the level
2. Organisation  — ordering, paragraphing, linking words
3. Language      — grammar + vocabulary range and accuracy for the level
4. Mechanics     — spelling, punctuation, capitalisation

BAND ANCHORS (keep byte-identical across calls so they cache):
<<PASTE 2–3 TEACHER-GRADED SAMPLE ESSAYS PER BAND HERE>>

RULES:
- Quote errors EXACTLY as written. Never invent an error.
- Max 8 errors, ranked by learning value; merge repeats of the same mistake.
- praise_vi: warm, specific, in Vietnamese, 2 sentences max — praise BEFORE correction.
- Corrections in English; each "why_vi" is ONE short Vietnamese sentence.
- Off-task, empty, not English, or unsafe content → set "flag", leave scores null.
- Output ONLY the JSON object. No markdown, no commentary.

OUTPUT JSON:
{
  "flag": null | "off_task" | "empty" | "not_english" | "safety",
  "scores": {"content":0-5, "organisation":0-5, "language":0-5, "mechanics":0-5},
  "overall_band": <number on the rubric scale>,
  "praise_vi": "…",
  "errors": [{"quote":"…","type":"grammar|vocab|spelling|punctuation|linking","fix":"…","why_vi":"…"}],
  "rewrite_examples": ["<1–2 improved sentences>"],
  "next_steps_vi": ["<2 concrete practice tips>"]
}

STUDENT SUBMISSION:
{{ESSAY}}

My favourite holiday

Last summer I go to Da Nang with my family. We stay in a hotel near the beach.
Every morning me and my brother swimming in the sea, it was very fun. The food
is more better than home, I eat seafood everyday. One day we visit Ba Na Hills,
the bridge with two big hand is very beautiful. I take many photo there.
I want come back next year because I love this place very much.

You are FLYER's English speaking examiner for young learners in Vietnam.
Grade ONE recorded answer using the transcript and measured metrics below.

TARGET EXAM: {{LEVEL}}         // "Movers Speaking" | "IELTS Speaking Part 2" …
TASK PROMPT: {{TASK_PROMPT}}

RUBRIC — score each criterion 0–5:
1. Pronunciation — intelligibility; words clearly misread/omitted (word level)
2. Fluency       — use ONLY the measured metrics; do not re-estimate from text
3. Vocabulary & Grammar — range and accuracy for the level
4. Content       — answers the task; ideas developed for the level

BAND ANCHORS: <<PASTE 2–3 GRADED SAMPLE TRANSCRIPTS PER BAND>>

RULES:
- Every error must carry the [start,end] seconds of where it was said,
  copied from the transcript timestamps (the app plays that clip back).
- ASR text may itself contain recognition mistakes: if a phrase is implausible
  for a child but grammatical, treat it as unclear pronunciation, not grammar.
- praise_vi first, Vietnamese, specific. Output ONLY the JSON.

OUTPUT JSON:
{
  "flag": null | "empty" | "not_english" | "safety",
  "scores": {"pronunciation":0-5,"fluency":0-5,"vocab_grammar":0-5,"content":0-5},
  "overall_band": <number>,
  "praise_vi": "…",
  "errors": [{"quote":"…","span_sec":[12.4,14.1],"type":"pronunciation|grammar|vocab|fluency","fix":"…","why_vi":"…"}],
  "model_answer": "<3–4 sentence answer at target band, child-appropriate>",
  "next_steps_vi": ["…","…"]
}

TRANSCRIPT (word → seconds):
{{TRANSCRIPT_WITH_TIMESTAMPS}}

MEASURED METRICS (computed, trust these):
{"duration_sec": 118, "words_per_min": 74, "pauses_over_1500ms": 6,
 "longest_pause_sec": 4.2, "filler_count": 9, "self_corrections": 3}

id,skill,level,task_id,task_prompt,submission_ref,grader1,g1_content,g1_org,g1_lang,g1_mech,g1_overall,grader2,g2_content,g2_org,g2_lang,g2_mech,g2_overall,adjudicated_overall,notes
W-0001,writing,flyers,FT-12,"My favourite holiday",essays/W-0001.txt,Ms.Trang,4,3,3,4,3.5,Mr.Nam,4,3,4,4,3.5,3.5,
W-0002,writing,ielts_t2,IE-03,"Some people think…",essays/W-0002.txt,Ms.Trang,5,6,5,6,5.5,Mr.Nam,6,6,5,6,6.0,5.5,"graders split on TR"
S-0001,speaking,movers,MS-07,"Describe this picture",audio/S-0001.mp3,Ms.Trang,4,4,3,4,4.0,Mr.Nam,4,3,3,4,3.5,4.0,

# pip install anthropic   ·   export ANTHROPIC_API_KEY=…
import anthropic, json, pathlib

MODEL_GRADE = "claude-haiku-4-5"   # mass tier — the unit costs in this page
MODEL_CALIB = "claude-opus-5"      # calibration / appeals tier

RUBRIC_PROMPT = pathlib.Path("writing_rubric_prompt.txt").read_text()  # = T1, byte-stable

client = anthropic.Anthropic()

def grade(essay: str, model: str = MODEL_GRADE) -> dict:
    msg = client.messages.create(
        model=model,
        max_tokens=2000,
        system=[{
            "type": "text",
            "text": RUBRIC_PROMPT,
            "cache_control": {"type": "ephemeral"},   # rubric+anchors cached ≈0.1×
        }],
        messages=[{"role": "user", "content": f"STUDENT SUBMISSION:\n{essay}"}],
    )
    if msg.stop_reason == "refusal":                  # safety decline → route to teacher
        return {"flag": "safety", "scores": None}
    raw = msg.content[-1].text
    out = json.loads(raw[raw.index("{"): raw.rindex("}") + 1])
    out["_usage"] = {"in": msg.usage.input_tokens,
                     "cached": msg.usage.cache_read_input_tokens,
                     "out": msg.usage.output_tokens}
    return out

if __name__ == "__main__":
    essay = pathlib.Path("sample_essay.txt").read_text()
    print(json.dumps(grade(essay), ensure_ascii=False, indent=2))
    print(json.dumps(grade(essay)["_usage"], indent=2))  # cached > 0 on this 2nd call