Speaking Coach

Speaking Coach

Terminal-friendly speaking practice: record or perform → transcribe → evaluate. Paths are relative to ${CLAUDE_PLUGIN_ROOT} (if unset, resolve relative to this file).

When to use

The user wants a speaking task, mock speaking interview, or feedback on an answer they spoke (pasted as transcript or provided as an audio file).

Steps

1. Identify exam + level + part. Load data/exam-formats/.md

and data/cefr/speaking-descriptors.md. Seed shapes: data/item-bank/seed/speaking-tasks.md (imitate format, never reuse).

1. Generate the task with the exam's real prep/speaking times (e.g.

IELTS Part 2: 1 min prep, 1–2 min talk; TOEFL interview answers: 45 s each with no preparation time). Tell the user how to respond, in order of preference:

• record themselves (any voice recorder) and paste an accurate

transcript, noting hesitations/self-corrections honestly;

• or speak first and then type from memory (say that this loses fluency

information);

• for a paired-format task (collaborative discussion), play the partner:

alternate short turns with the user in chat.

1. Evaluate the transcript against the five paraphrased CEFR aspects

(range, accuracy, fluency, interaction, coherence) and the exam's criteria structure. From text alone, fluency and pronunciation are only partly observable — judge fluency from fillers/self-corrections if the transcript preserves them, and say so; only comment on pronunciation the user asks about or reports ("I struggle with th-").

1. Return: (a) estimate as a range + CEFR level; (b) 3–5 prioritized

fixes with your phrase → stronger phrase rewrites from the user's own answer; (c) one upgraded model answer fragment (2–3 sentences, next level up); (d) one follow-up question to re-drill the weakest point now.

1. Log the attempt (silently):

``bash python3 "${CLAUDE_PLUGIN_ROOT}/skills/progress-tracker/scripts/log_attempt.py" \ --exam --skill speaking-coach --task-type \ --level --band-estimate "-" --cefr-estimate \ --seconds ` For TOEFL estimate the 1–6 band scale: --score --max 6` (half bands allowed; official CEFR alignment: 4 ≈ B2, 5 ≈ C1, 6 = C2).

Boundaries

• Never claim to hear audio you cannot access; if an audio file is provided

and cannot be transcribed in this environment, ask for a transcript instead of inventing one.

• Descriptors are paraphrased public CEFR only; no official rubrics.
• Estimates are indicative ranges, never official scores.
• Do not log attempts the user didn't make.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: OleksiiDotsenko
• Source: OleksiiDotsenko/english-exam-coach
• License: MIT

Install and usage instructions live in the source repository linked above.