메뉴
Oral-reading-fluency assessment (EGRA/ASER-style) — the passage→record→transcribe→score pipeline, WCPM benchmarks, multilingual rules, and debugging.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Classroom-observation coaching pipeline — pluggable scoring frameworks (OECD/HOTS/TEACH/FICO), the async queue worker, lesson-plan integration, and session debugging.
Trace any feature end to end — webhook → handler → service → queue → worker → DB → storage → delivery — to find where a user's request actually broke. The map that points you at the right service, worker, table, and skill for each feature.
How the bot logs — structured single-line JSON to the console, correlation IDs that thread a request across handler/queue/worker, semantic events, and an optional external log backend. Use when adding logging or wiring observability.
Pre-merge defensive checks for recurring, predictable bug classes — run before merging code that (a) emits new interactive button/list/Flow IDs, (b) edits or sends a Flow, (c) writes new enum/status values, (d) chains Supabase update+filter+select, (e) sends to a dormant user, or (f) tracks files toxic to a fresh clone. Each class has a concrete grep/SQL pre-flight and the failure mode it prevents.
A/B testing via a Thompson-sampling multi-armed bandit. DB-driven tests (ab_tests / ab_test_variants / ab_test_events) that learn the best-performing variant online. Use when adding or measuring a message/content experiment.
Query the bot's Postgres database with a read-only role. Use when analysing user metrics, debugging DB state, or exploring the schema. Column names: phone_number (not phone), first_name (not name) in the users table.
| name | reading-assessment |
| description | Oral-reading-fluency assessment (EGRA/ASER-style) — the passage→record→transcribe→score pipeline, WCPM benchmarks, multilingual rules, and debugging. |
Up: .claude/CLAUDE.md (config & skills router) · See also: coaching, debugging
A teacher runs a student through a short oral-reading assessment: the bot generates a passage, the student reads it aloud, and the bot transcribes the audio, scores fluency against grade benchmarks, and returns a report with voice feedback.
passage-generation, transcription, pronunciation, fluency, analysis, comprehension, report, voice-feedback./reading command routed through bot/shared/handlers/text-message.handler.js.EGRA/ASER-style Oral Reading Fluency (ORF) assessment. Primary metric is WCPM (Words Correct Per Minute); alphabet assessments use LCPM (Letters Correct Per Minute). Passage types scale by difficulty: letters → words → sentences → paragraphs → story.
compareToBenchmarks → the check_benchmark_status RPC) → report + voice feedback.grade_levelgrade_level is the difficulty proxy (0 = letters … up to story) and drives both passage generation and the
benchmark lookup. The benchmark RPC check_benchmark_status(p_wcpm, p_grade, p_language, p_is_l2) returns
(benchmark_min, benchmark_max, on_track, percentile_rank), reading from the wcpm_percentiles table
(25th percentile = min, 75th = max; season auto-derived from the current month).
Once the teacher picks the assessment language, every bot-generated string for the rest of that
assessment must be in that language — not the user's stored preferred_language. The welcome message
before the pick still uses the user's language. Pin each prompt with an explicit
MUST be written entirely in language code "${lang}", because the model otherwise drifts back to the
default. (Comprehension answers are preserved verbatim for non-English; the report renderer handles RTL +
Nastaliq natively.)
| Language | Pronunciation scoring | Benchmarks |
|---|---|---|
| English | Full (provider-based) | Hasbrouck-Tindal ORF norms |
| Urdu | LLM fallback (no provider scoring) | L2-adjusted (~70% of EN; recalibrate after enough production data) |
| Arabic | LLM fallback, RTL | L2-adjusted |
| Others | LLM fallback | extrapolated |
RTL languages (Urdu/Arabic) render grids right-to-left with reversed letter order and direction indicators.
| Comprehension score | Label |
|---|---|
| ≥ 80% | On Track |
| 60–79% | Approaching Benchmark |
| < 60% | Needs Support |
Performance tiers off the fluency percentile: ≥75 Excellent · on-track & ≥50 Proficient · on-track Developing · else Emerging — applied identically in the HTML template and the legacy renderer so both look consistent.
| Symptom | Cause | Fix |
|---|---|---|
| Voice feedback has nonsensical corrections | Transcript not cleaned (diarization labels left in) | Clean the transcript before word alignment. |
| Only one comprehension question asked | Voice routing priority wrong (reading handler before comprehension) | Check routing order in voice-message.handler.js. |
| Repetitive passages | No diversity controls in the prompt | Use the theme/name/avoid arrays in passage generation. |
| Wrong benchmark | Grade/language/L2 args mismatched | Confirm the check_benchmark_status args; trace the correlation id — see debugging. |
const correctWords = passageWords.filter((w, i) =>
transcriptWords[i]?.toLowerCase() === w.toLowerCase()).length;
const wcpm = (correctWords / (durationSeconds / 60)).toFixed(1);