ponytaillisted

# Ponytail Most useful tools are approximations. That is fine. An approximation **presented as a fact** is not: the reader trusts the output, it is wrong in a way nobody wrote down, and the trust is spent before anyone notices. The rule is `rules/ponytail.md`. This is how to apply it to a diff. ## 1. Find what owes a line ```sh .claude/tools/ponytail.sh # whole tree .claude/tools/ponytail.sh path/to/changed.go ``` It flags a file that reads as best-effort and carries no `Ponytail:` marker. Read its own header first — it over-reports on hedge words in ordinary prose and under-reports badly on approximations written without any tell, so treat every row as a prompt to look, never a verdict. Then check what the tool cannot see in the change itself: - Parsing source with a regex instead of an AST. - A sample, a `head -n`, a `LIMIT`, a partial scan standing in for a full one. - Existence used as a proxy for a property — a test file's *name* standing in for coverage. - A timeout, a retry bound, a cache TTL: each one is a guess about how long is too long. - A number that is derived rather than measured — bytes÷4 for tokens, a ratio, an extrapolation. - A default chosen because it worked on your machine. ## 2. Write the line Not "this is regex-based" — that tells the reader nothing. State **what fails, in which direction, and what to do instead**: ```text # Ponytail: matches the import line, so a conditional or dynamic import is missed. # It under-reports, so