learn-coach

name	learn-coach
description	Interactive coach for `python trainer.py learn`. Asks a few questions about your stack and the words your voice model mis-hears, seeds the miner, reviews risky suggestions before writing, and runs an independent agent to brainstorm improvements. Use when building or refreshing a SuperWhisper config from your own dictation history.

learn-coach

A guided wrapper around python trainer.py learn. It turns a raw history-mining run into a short coached session: a few questions steer the miner toward your stack, you review anything risky before it's written, and an independent agent proposes improvements each run.

[!NOTE] This skill is optional and Claude-Code-specific. The underlying command works without it: python trainer.py learn --hints <file> does the same mining. The skill just collects the hints for you and adds the review + brainstorm steps. Step 5 (brainstorm) needs an agent-capable runtime and is opt-in.

Step 1 — Ask the feeder questions (offer Skip)

Ask the operator the questions below. Make Skip a first-class option — say explicitly: "Reply skip to run with defaults and no questions." If they skip (or don't answer), jump straight to Step 3 and run with no hints.

1. Systems — tools, languages, frameworks, products you dictate about. Seeds known-term matching (e.g. Cube, dbt, WP Fusion).
2. Mis-hears — words the voice model reliably gets wrong, as heard → meant (e.g. cloud code → Claude Code, active campaign → ActiveCampaign).
3. Leave alone — look-alike words it should NOT "correct". (You really do say "cloud code" about infrastructure.) These become the suppress list and stop false fixes before they happen.
4. Pointers — client / product / proper names worth forcing into vocabulary.

Step 2 — Write the hints file

From the answers, write user_configs/learn_hints.yaml (git-ignored). Omit any empty key. Full schema: data/learn_hints.example.yaml.

systems: [Cube, dbt, WP Fusion]
misspellings:
  - from: "cloud code"
    to: "Claude Code"
vocabulary: [Acme Corp]
suppress: [cloud code]

Step 3 — Run learn

python trainer.py learn --hints user_configs/learn_hints.yaml
# …or, if the operator skipped Step 1:
python trainer.py learn

Add --source <path> to mine an aggregated cross-device folder (see docs/IMPORTING_HISTORY.md).

Step 4 — Review false positives

Open the generated user_configs/from_history.yaml and flag any risky replacement for the operator:

• a common English word on the from side (could rewrite ordinary speech)
• any from → to pair the operator might not actually want

Show the flagged rules; let the operator drop any. Apply their choices by editing the file, or by adding the rule's from to suppress and re-running Step 3. Leave the safe rules (case-fixes, multi-word names) in place.

Step 5 — Independent improvement brainstorm (opt-in)

Spawn ONE sub-agent with no context from this conversation to propose miner improvements independently, so each run gets a fresh outside view.

• SWARM CONFIG: count=1 model=opus effort=high est_cost_tier=high — independent design synthesis. Skip if the operator declines or no agent runtime is available.
• If a local deep-planning command (e.g. /deep-plan) is available you may use it; otherwise dispatch a plain Opus agent. Do not hard-depend on any one command.
• Feed it the newest file in user_configs/brainstorms/ (if any) as "already proposed — return ONLY net-new ideas," so runs converge instead of repeating.
• Brief it on: the miner's approach (corpus mining + known-terms fuzzy matching + hints), the meta.json fields available (result, rawResult, segments, modeName, modelName, datetime, duration), and ask for concrete, stdlib-preferred precision/recall improvements ranked by value/effort, each with its false-positive risk.
• Save the output to user_configs/brainstorms/<YYYY-MM-DD-HHMM>.md (git-ignored) and summarize the top 2–3 net-new ideas for the operator.

Step 6 — Apply (optional)

If the operator approves the reviewed config:

python trainer.py train --config user_configs/from_history.yaml

Quit SuperWhisper first so it can't overwrite the write from memory; a timestamped backup is created automatically, and train skips rules already in your settings.