regex-builder

Regex Builder & Explainer Skill

Produce correct, readable regular expressions — and explain them so the user actually understands what they're shipping.

Working from a brief

Infer the regex flavor (JavaScript/PCRE/Python/Go) from context; if unstated, default to one and say so (assumed — confirm). Always deliver a working pattern and tests even from a loose description. Never leave placeholders.

Required Inputs

• What should match and what should NOT — 3+ positive examples and, critically, 2+ near-miss negatives (the strings that look matchable but must be rejected). The negatives are where every regex bug lives.
• The engine/flavor (JavaScript, PCRE, Python re, RE2, grep -E…) — anchors, lookbehind, and Unicode behaviour differ enough to break portability silently.
• Where it runs — validation, extraction, or replacement changes how greedy the pattern should be.

Two modes

• Build: the user describes what to match → produce the regex.
• Explain: the user pastes a regex → break it down.

Detect which from the input.

Output Structure

Pattern

The regex in a code block, plus the flavor and any flags (e.g. i, g, m) and why.

Breakdown

A token-by-token table or list: each part of the pattern and what it matches.

| Token | Matches |

|-------|---------|

| ^ | start of string |

| … | … |

Test cases

• ✅ Matches: 3–5 strings it should match
• ❌ Rejects: 3–5 strings it should not match (include the tricky near-misses)

Notes

Edge cases, catastrophic-backtracking risks, anchoring, Unicode, and a simpler alternative if the regex is getting unwieldy (sometimes "don't use regex" is the right answer — say so).

Quality Checks

• [ ] The pattern actually passes the listed "matches" and rejects the "rejects"
• [ ] Flavor and flags are stated
• [ ] The breakdown covers every token, not just the interesting ones
• [ ] Edge cases / backtracking risks are flagged

Anti-Patterns

• [ ] Do not give a regex with no test cases — always prove it
• [ ] Do not ignore the flavor — \d, lookbehind, and named groups differ across engines
• [ ] Do not produce an unreadable one-liner when a commented/verbose version or a non-regex approach is clearer
• [ ] Do not silently assume anchoring — state whether it matches the whole string or a substring