docs: grade the letter-count examples with an exact match

[Mike-7777777]

Problem

The letter-count grader in the README checks str(word.count(letter)) in answer, which is a substring match. For strawberry (3 r's), the answers "13", "33" and "2 or 3" all get a reward of 1.0.

The same check appears in the quickstart, docs/skill.md, and twice in the "Creating Environments" guide. The guide's version is also the snippet shown in the "What is a good task?" post.

Change

Compare the stripped answer instead, the same check environments/blank/env.py already uses:

yield 1.0 if (answer or "").strip() == str(word.count(letter)) else 0.0

The README and quickstart prompts already say "Reply with just the number." The docs/skill.md and guide prompts didn't, so I added the same sentence there. Otherwise a correct answer like "3 times" would score 0 with nothing in the prompt asking for a bare number.

Trade-off: "3." and "The answer is 3." now score 0, which matches what the prompt asks for. If you'd rather be lenient, hud.graders.numeric_match(answer, word.count(letter)) checks the first number in the answer. It also rejects "13", at the cost of one import in the snippet. Happy to switch.

Check

• Ran the README snippet with the local hud install and drove the generator directly, no model calls: "3" scores 1.0; "13", "2 or 3" and None score 0.0; "2" for blueberry scores 1.0.
• Ran each edited grader line from the four files on "3", " 3\n", "3.", "The answer is 3.", "13", "33", "2 or 3", "not 2, it's 3", "three", "" and None. Only "3" and " 3\n" score 1.0.
• ruff format --check . and ruff check . pass.

Side note, not changed here: the quickstart says hud init my-env --template blank creates a tasks.py containing count_letter(word, letter). The blank template actually ships count(sentence, letter) in env.py, and its tasks.py imports it.

---

Note

Low Risk
Docs-only changes to example graders and prompts; no runtime or SDK behavior is modified.

Overview
Documentation examples for the letter-count task now grade with an exact match on the stripped answer ((answer or "").strip() == str(count)) instead of a substring check (str(count) in answer), so false positives like "13" or "2 or 3" no longer receive full reward.

Where prompts did not already ask for a bare number, docs/skill.md and creating-an-environment.mdx (both snippets) add "Reply with just the number." so grading stays aligned with what users are told to return. README, quickstart, and the environment guide’s case-insensitive count template are updated the same way.

^{Reviewed by Cursor Bugbot for commit 18d2bc3. Bugbot is set up for automated code reviews on this repo. Configure here.}