exam-router

Exam Router

Entry point and orchestrator. Resolves which exam, which level, which section, then delegates. Does not evaluate work itself.

All plugin file paths below are relative to ${CLAUDE_PLUGIN_ROOT} (the plugin's install directory; if the variable is unset, resolve paths relative to this SKILL.md file, two directories up).

Supported exams

Steps

1. Resolve the request to exam id + CEFR level + section. Infer where possible ("FCE writing" → cefr-b2, B2, writing). If the exam is ambiguous AND matters for the task, ask one short question; otherwise default sensibly and say what you assumed. If the user names an English exam that is not one of the seven supported ids, do not silently map it onto a supported one — handle it per Boundaries.
2. Load ${CLAUDE_PLUGIN_ROOT}/data/exam-formats/<exam-id>.md. Answer format questions directly from it — never from memory.
3. Hand off by section: writing → writing-evaluator, speaking → speaking-coach, reading/use of english → reading-use-of-english, listening → listening-trainer, vocabulary → vocabulary-builder, study plan / exam date → study-planner, results → progress-tracker.
4. Level diagnostic (when asked "what's my level" or via /assess-level):
   • Items: 8 use-of-english items, 2 per level B1–C2 (mix open cloze, multiple-choice cloze, one inference item), presented easiest-first. Stop rule: two consecutive misses at a level caps the estimate there — do not keep probing above it.
   • Writing sample (80–120 words) is the primary evidence: place it against ${CLAUDE_PLUGIN_ROOT}/data/cefr/calibration-anchors.md per criterion; the items refine, they don't outvote it.
   • Optionally add 3 can-do self-checks from ${CLAUDE_PLUGIN_ROOT}/data/cefr/can-do-statements.md.
   • Report a CEFR range (e.g. "B2, approaching C1"; "A2, approaching B1" for lower starters), state it is indicative, and state what was not measured (listening and speaking are not in this probe).
   • Log it exactly once. This Step 4 call is the single source of the level-diagnostic record; when the diagnostic is reached via /assess-level (or any other entry point), do not log the same diagnostic a second time. Use every required field, with --cefr-estimate (that is what feeds cross-exam trends — a bare --band-estimate "B2-C1" has no numbers and normalizes to nothing):

     python3 "${CLAUDE_PLUGIN_ROOT}/skills/progress-tracker/scripts/log_attempt.py" \
       --exam <target exam or cefr-XX> --skill exam-router \
       --task-type level-diagnostic --level <nearest anchor> \
       --cefr-estimate <estimated level> --score <items correct> --max 8 \
       --seconds <measured>
     

     (Windows without python3: run the same line with python or py.) The level-diagnostic row is reference-only — a level-establishing probe, not a drillable task — so never surface it as a "weakest task type" to drill.
5. Suggest a next step grounded in the user's goal (target exam, weakest section, or /daily-drill).