Coach

What it does: shows the numbers behind your sessions that drive cost and quality, checks your habits against a normal software lifecycle (SDLC), and teaches one architect-level concept at a time, tied to what you just did. The aim is to carry vibe-coding skill into real engineering work.

Where you see it:

• C in the dock (or /coach) toggles the panel:

  ● Cache    92%   ━━━━━━━━━━━  reusing context well
  ● Context  41%   ━━━━━━━━━━━  room to spare
  ● Cost     $3.20             $0.27 per turn
  Habits   ✓ Plan  ✓ Branch  ✗ Tests  ✓ Commits  · PR
  Do next  Run the tests: 8 edits since the last run
  Tokens   120k fresh  4.1M cached  300k cache-write  52k out
  Model    claude-opus-5-5  effort high  subagents 4 runs · 31% of tokens
  Lesson   The test pyramid
           Many fast unit tests at the base, fewer integration tests…
  

  • Colors: green is good, yellow means watch it, red means act. In the token row, cached tokens are green (cheap), cache-writes yellow and output cyan.
  • More adds the lesson's "try next" and "at work" notes (with Another lesson) and a Claude Code pro tip (with Next tip and links).

• Hover C in the dock for a violet card with the cache and context bars, cost, habits and "do next".

• One tip after a turn, only when a rule fires, as a dim 🎓 Coach · … row in the chat:

  • committed straight to main
  • cache hit under 50% on a big turn (with the likely cause)
  • context over 75% full
  • 5+ edits since tests last ran
  • a turn costing over ~$2
  • 10+ minutes of building with no plan
  • subagents using over half the tokens

  The same tip waits at least 6 turns before repeating.

Glossary:

• Prompt cache: the app re-sends the whole conversation every turn. A cache hit means that prefix was reused and billed at ~10% of fresh input. Hits break on model switches, CLAUDE.md or memory edits, /compact, or idle gaps (the cache expires).
• Context fill: how much of the model's window the conversation uses. Fuller means slower, pricier, and lower quality near the limit.
• Cost: API prices. On a subscription this is the value you used, not a bill, but it tracks how fast you burn your 5-hour window.

How it works: everything comes from what the app already reports (token usage per turn, session cost and context, tool calls such as edits, tests, commits and gh pr create). No model calls.

Cost: none.

Data saved: none on disk. Stats live for the session.