/drill

User Input

$ARGUMENTS

Workflow

Step 1: Parse arguments

Flag	Default	Meaning
--category <name>	(none)	Restrict drill to one category: spelling / grammar / punctuation / word-choice / article / preposition
--rounds N	3	How many drill sentences to generate

Step 2: Load all-time mistake patterns

Use the shared JSONL parser pattern (see commands/shared/jsonl-parser.md) to read correction history.

node -e "
  const { periodStats } = await import('${CLAUDE_PLUGIN_ROOT}/scripts/lib/stats.mjs');
  console.log(JSON.stringify(periodStats(365)));
"

If the script fails, read JSONL files directly via Glob + Read following the fallback pattern in commands/shared/jsonl-parser.md.

If the user has fewer than 5 corrections total across all time, respond: "Not enough history yet to drill. Use Claude Code for a few days first, then try /claude-english-buddy:drill again." and STOP.

Step 3: Identify top-3 categories

Group all corrections by category (spelling, grammar, punctuation, word-choice, article, preposition). Rank by total count. Take the top 3.

If --category was passed, use that single category instead.

Step 4: Generate drill sentences

For each of the N rounds (default 3):

1. Pick a category from the top-3 (round-robin across rounds if 3; otherwise random from filtered set).
2. Pick a specific pattern the user has made inside that category (e.g. its got → it has).
3. Generate ONE new sentence that uses the wrong form of that pattern. The sentence should:
   • Be a plausible developer-context sentence (commit message, PR description, doc blurb, error message).
   • Contain exactly ONE intentional error matching the target pattern.
   • Not repeat a sentence the user has seen before (if possible).
4. Present the sentence and ask the user to correct it.

Step 5: Present drill

# Drill Round {1 of N} — Category: {category}

Your recurring pattern: **{original → corrected}** (seen {count} times)

## The Sentence

> {generated sentence with the intentional error}

Rewrite this sentence to fix the error. Reply with the corrected version; I'll grade your answer and show the rule.

After the user replies with a corrected sentence, grade it:

## Grading — Round {N}

| | Target pattern | Your correction | Verdict |
|-|----------------|-----------------|---------|
| Original | {sentence with error} | — | — |
| Expected | {sentence with fix} | — | — |
| You wrote | — | {user's answer} | {Correct / Partial / Incorrect} |

### Rule

{One-sentence explanation of the underlying rule — pulled from grammar-fundamentals, punctuation-rules, or common-non-native-mistakes as appropriate.}

{If user got it correct: brief reinforcement.}
{If partial: name the thing they missed.}
{If incorrect: state the rule and show the minimal edit.}

After N rounds, emit a summary:

## Drill Summary

| Round | Category | Verdict |
|-------|----------|---------|
| 1 | ... | Correct |
| 2 | ... | Partial |
| 3 | ... | Incorrect |

**Score**: {n/N correct, n/N partial, n/N incorrect}
**Weakest pattern today**: {pattern}
**One tip**: {actionable guideline from the underlying skill}

Notes

• This is a learning tool. Be encouraging. If the user gets all answers wrong, do not pile on; focus on one pattern to work on.
• Do not log drill attempts to the JSONL correction history — they are practice, not real usage.
• Grading uses string comparison for the TARGET pattern only. If the user makes OTHER edits (stylistic rewrites), accept them as long as the target pattern is correctly fixed.
• If --category names a category the user has zero mistakes in, fall back to their actual top category and note the fallback in the first round header.