reSkill

A terminal companion for Claude Code. While Claude thinks for 23 seconds, reSkill hands you a quiz about the exact code it's working on.

Zero network. No Claude slowdown. Local state, single flag file. MIT, alpha. Landing: https://amitsubhash.github.io/reskill/.

Claude is thinking...                   reSkill   Q1
                                        ╭── think about this ──╮
  · Reading 3 files                     │ Most Pythonic way to │
  · Analyzing scheduler.py              │ read a UTF-8 JSON?   │
                                        │                      │
  Actioning... 16s                      │  1) open(...)        │
                                        │  2) Path(...).read_  │
                                        │     text(utf-8)  ✓   │
                                        │  3) codecs.open()    │
                                        │  4) io.open()        │
                                        ╰──────────────────────╯

Install

# 1. install the package  (the PyPI name is reskill-claude; the CLI binary is `reskill`)
pip install reskill-claude

# 2. see it work, no settings touched
reskill demo

# 3. wire it into Claude Code (reversible; backs up your settings.json)
reskill install
reskill doctor

Prefer editable? git clone https://github.com/AmitSubhash/reskill.git && cd reskill && pip install -e . Or install straight from GitHub: pip install git+https://github.com/AmitSubhash/reskill.git.

Use

Live during Claude sessions (the main one):

reskill claude   # opens claude + a quiz pane alongside

If you have tmux: splits the current window. If not: pops a second Terminal.app or iTerm2 window. Click the quiz pane (or Ctrl+B →) to focus it, press 1 to 4 to answer.

Anytime, anywhere:

reskill next                    # one context-matched quiz, right now
reskill next --concept torch    # target a specific topic
reskill session                 # commit-driven deck (last 7 days)
reskill review                  # drill your recently-missed questions
reskill topics                  # learning map: every concept + mastery
reskill doctor                  # diagnose anything that feels off
reskill status                  # one-liner: 0 mastered, 3/5 today, 5 streak
reskill stats                   # level, XP, per-concept mastery
reskill streak                  # 12-week heatmap of your answered days

How it works

• Hooks fire on UserPromptSubmit, PreToolUse, PostToolUse, Stop to toggle a flag file at ~/.reskill/state/thinking.
• Quiz pane watches the flag (or falls back to transcript-mtime polling if hooks aren't installed) and serves a question whenever Claude is mid-thought.
• Scheduler picks questions matched to what Claude is writing. Tiers: live transcript > recent git commits > cumulative cache. Within a tier: SM-2 overdue > new > not-due. Across concepts: interleaved within confusable clusters (Rohrer & Taylor 2007). Within a concept: format diversity (Roediger & Karpicke 2006). Targets a 15% error rate per Wilson 2019.
• Pacing gate rate-limits with sensible defaults, tunable via RESKILL_* env vars.
• In-session re-queue pushes wrong answers back 3 items later (Butler & Roediger 2008).
• Hypercorrection cue: wrong answers given in under 5 seconds get a sticky flag. Research says these stick hardest.

Controls

Themes

export RESKILL_THEME=everforest  # default on truecolor terminals
export RESKILL_THEME=mono        # BOLD/DIM only
export NO_COLOR=1                # standard, also forces mono

Auto-falls-back to mono when COLORTERM is not truecolor or 24bit (Apple Terminal.app, many ssh contexts).

Tuning

export RESKILL_MIN_GAP=10                # seconds between quizzes
export RESKILL_MAX_PER_HOUR=20
export RESKILL_MAX_PER_DAY=60
export RESKILL_THINKING_DEBOUNCE=3       # skip first N seconds
export RESKILL_SAME_CONCEPT_COOLDOWN=90

What's inside

• 74 hand-written questions across 50 concepts
• Python (async, typing, numpy / pandas / torch, stdlib gotchas, packaging, testing) plus shell, git, SQL, React and TypeScript essentials
• 8 question formats: output, bug, tradeoff, scenario, why, gotcha, refactor, cloze
• LLM-generated questions from your live transcript or real commit diffs via reskill gen --live

Uninstall

reskill uninstall   # removes hooks + statusline, keeps a backup

Your existing settings.json hooks and statusLine are preserved. We only touch entries we own.

Status

Alpha. Data model stable, scheduler is evidence-based. The template bank is hand-written; LLM-generated questions from novel diffs are the next major milestone.

License

MIT. See LICENSE.

Credits

Built by Amit Subhash (@AmitSubhash). Scheduler grounded in the spaced-repetition literature: Bjork & Bjork 2011 (desirable difficulties), Rohrer & Taylor 2007 (interleaving), Butler & Roediger 2008 (delayed feedback), Wilson et al. 2019 (85% rule), Butterfield & Metcalfe 2001 (hypercorrection). Mistakes are mine.