Claude Coach

Six of them

The tools

The authoritative list is whatever tools/list returns — this page is the version for humans.

review_prompt

Reads a prompt and reports what is missing, with a score out of 100 and a specific fix beside each finding: no output format, no audience, no constraints, vague words, code with no language named, a document question with no permission to say “not stated”.

improve_prompt

Rewrites a rough prompt into context / task / constraints / output format. What you did not supply comes back as a TODO rather than an invented constraint — a guessed requirement reads exactly like a decided one.

guidance

Eleven practices, by id or by search: being explicit, structuring a long prompt, examples, output format, when thinking pays, long context, tool design, agents, Claude Code, iterating with an eval set, and keeping answers honest.

recommend_model

Recommends a tier — most capable, balanced, fast — for a described task, with how to verify the choice. Tiers rather than model ids, on purpose: ids change and this advice does not get redeployed when they do.

save_lesson

Keeps something that worked — a phrasing, a workflow, a setting — in the workspace's database, so it outlives the conversation it was learned in.

list_lessons

Hands those lessons back, newest first, optionally filtered by tag. Worth calling at the start of a coaching conversation.

What it will not do

It will not answer your prompt for you, call another model, reach the network, or store anything you did not ask it to store. The tools that write take one field and a tag; everything else is a pure function of what you passed in.