Objective drills: generate → answer → score → explain → log. Paths are
relative to ${CLAUDE_PLUGIN_ROOT} (if unset, resolve relative to the plugin
root — the directory two levels above this file, i.e. ../../ from here).
The user asks to drill any reading or use-of-english task type, or asks why
a reading answer is what it is.
-
Identify exam + level + task type (ask once if unclear; /daily-drill
picks the weakest type from the progress log instead). Load
data/exam-formats/<exam-id>.md for the exact shape (items per task,
options per question) and data/cefr/reading-descriptors.md +
data/cefr/reading-calibration-anchors.md to calibrate text difficulty.
Seed shapes: data/item-bank/seed/reading-use-of-english-items.md.
Check the requested task type is actually in the chosen exam (per its
format file): key word transformation and word formation exist only in
B2–C2, cross-text matching only in C1, and so on. If the user asks for a
task their exam does not contain, say so plainly and offer the nearest
valid task or name the exam(s) that do include it — never generate an
off-format item.
True/False/Not Given vs Yes/No/Not Given: T/F/NG statements are about
FACTUAL information in the text; Y/N/NG statements are about the WRITER'S
views/claims (so the passage must carry opinions). Use the correct labels
for each.
-
Generate an ORIGINAL passage and items matching the format exactly:
right number of items, right option count, plausible distractors, one
defensibly correct answer each. Calibrate lexis and syntax to the CEFR
level. Give a realistic time budget (e.g. ~1.3 min per use-of-english
item; proportional share of the section time for reading sets).
Match the passage to authentic length — this is as important as the
item count. Aim for the upper end of the target range and count your
words before presenting: left to instinct these passages come out ~30–40%
too short. Use the exam's figure from data/exam-formats/<exam-id>.md
("Passage lengths"); as a fallback:
- Gapped text: B2 ~500–600, C1 ~550–780, C2 ~700–800 words of base text.
- Long-text multiple choice / multiple matching: B2 ~500–700, C1 ~700–800,
C2 ~700–800 words of base text. IELTS Academic passages ~700–900 words
each; IELTS General Training sections vary — see
ielts-general.md. A
too-short passage is the most common failure — a C2 Part 6 gapped text must
be ~750 words, not ~450.
- Cloze / word formation: ~150–220 words. TOEFL academic passage ~200;
TOEFL daily-life texts 15–150.
- Gapped text always has ONE more option than gaps (B2/C1/C2) or three
more (B1 Part 4) — the surplus fits no gap and is a deliberate distractor.
Removed paragraphs run ~50–80 words each.
Verify the length mechanically, do not estimate it. Before presenting,
write the passage to a temp file and count it — self-estimated word counts
are exactly the failure this guards against:
tmp="$(mktemp)"; cat > "$tmp" <<'PASSAGE'
...passage text...
PASSAGE
wc -w "$tmp"; rm -f "$tmp"
If the count is below the target range, extend the passage and re-count
before showing it. Never present a passage you have not counted.
-
Withhold the key. Present only the passage, the items, and the time
budget. Ask the user to answer all items (e.g. "1 B, 2 A, …") and note
how long they took.
-
Score objectively — one mark per item, no partial credit unless the
exam gives it (key word transformations are marked out of 2 in the B2–C2
exams: award 1 for a half-correct split). For open-ended item types
(open cloze, sentence/summary completion, short answer, word formation),
credit ANY answer that is valid and correctly spelled — one that fits the
context and, for cloze, the visible letters and the word limit; the printed
key is one acceptable answer, not the only one. If the user answered only
some items: if they ran out of time or stopped early, log against the
number actually attempted (--max <attempted>), not the full set; if they
genuinely gave up, count blanks as wrong; if nothing was attempted, log
nothing. Never fabricate answers for unattempted items. Then explain EVERY item, right
or wrong: why the key is correct, why each tempting distractor fails, and
for Not Given items, exactly what the passage does not say.
-
Log the attempt (silently):
python3 "${CLAUDE_PLUGIN_ROOT}/skills/progress-tracker/scripts/log_attempt.py" \
--exam <exam-id> --skill reading-use-of-english --task-type <slug> \
--level <anchor> --score <n> --max <total> --seconds <time>
Then, for each item the user missed, log why it was missed — a
comprehension subtype (detail, inference, paraphrase, distractor,
location, instruction) with a --point a fresh item could re-test, e.g.
"not recognising a paraphrase of 'decline'" — and run queue.py sync.
The item number is not re-testable; the reason is. Full protocol:
../progress-tracker/references/the-loop.md.
Use the exact --task-type slug from data/task-types.md so attempts
aggregate across sessions. (On Windows, if python3 isn't found, use
python or py.)
For out-of-2 task types (key word transformation), --max is 2 × the
number of items scored and --score is the sum of the 0/1/2 marks — e.g. a
6-item KWT set is --max 12, a 4-item partial attempt is --max 8.
log_attempt.py rejects a --score above --max and writes nothing, so
an under-sized --max silently drops the attempt.
Then offer one of: re-drill the same type, step up a level, or switch to
the weakest type from the log.
