Skip to content

v1.2.1 — Evaluate-first discipline

Latest
Latest

Choose a tag to compare

View all tags
@Terryc21 Terryc21 released this 16 May 13:12
v1.2.1
fa5820f

First public release after a full skill-reviewer audit. The skill went from v1.0.0 → v1.2.1 in this batch through three review rounds and four version bumps — all surfaced in a single review session. v1.2.1 consolidates the four bumps into the first release tag.

Headline change

prompter now defaults to evaluation, not rewriting.

Before producing any rewrite, the skill asks: would rewriting this prompt for clarity, typo correction, or actionability meaningfully improve it? It rewrites only when there's a real improvement to make — typos to fix, ambiguous references to resolve, vague intent to specify, or missing scope guardrails. If the original is already clear and actionable, prompter proceeds without rewriting and emits a canonical "no rewrite needed" notice.

This discipline is now stated four times across SKILL.md (opening declaration, frontmatter description, Rule 1, Rewriting guidelines) and once in the README intro, so an LLM reading from any entry point gets the right default behavior.

What changed from the original

Activation & lifecycle

  • Frontmatter description now lists natural-language activation phrases ("rewrite my prompt," "fix the typos in this prompt," "make this prompt clearer"), so the Claude Code router can auto-invoke without users typing /skill prompter
  • Re-invocation behavior is now defined for every mode (Try once, Next N, This session, Add/Remove CLAUDE.md)
  • N-prompts countdown has an explicit state model — the status line **Prompter:** N rewrites remaining is the source of truth, with an operational caveat about summarizer behavior on long runs

CLAUDE.md add/detect/remove symmetry

  • "Add to CLAUDE.md" now detects an existing Prompt Rewriting section via regex (^## Prompt Rewriting\\b) with five inline match/no-match examples, and defaults to preserving local edits rather than overwriting them
  • "Remove from CLAUDE.md" has a five-step removal procedure with explicit handling for modified blocks
  • The CLAUDE.md template is now framed as a seed; users are expected to evolve it locally and the skill protects those evolutions

Skip-notice & parseability

  • Rule 7 (skip notice format) specifies four canonical phrasings with the **Prompter:** prefix
  • Downstream parsers can disambiguate via suffix regex (skip notice vs no-rewrite vs countdown vs expiration)

Examples & README

  • 20 worked examples covering typos, dangling references, stacked questions, hedging preservation, no-rewrite cases, and threat-model framing
  • Top-of-file Contents list pairs each example with a one-phrase pattern tag (diagnose-before-fix, resolve dangerous verb, preempt blast radius)
  • README "Why this exists" leads with evaluation framing; install section acknowledges plugin packaging is on the roadmap

Install

Manual install only — see README for global and project-specific paths.

git clone https://github.com/Terryc21/prompter.git
mkdir -p ~/.claude/skills && cp -r prompter ~/.claude/skills/

Commits in this release

  • `fa5820f` — fix: align entry-point text with v1.2.0 evaluate-first discipline (1.2.1)
  • `780808d` — feat: evaluate-first discipline rule + bump to 1.2.0
  • `e5dea63` — chore(version): bump to 1.1.0 for skill-reviewer fix waves
  • `c457250` — fix: apply 7 second-pass skill-reviewer findings
  • `ab13dd3` — fix: apply 12 skill-reviewer findings (deferred 1 intentional)
Assets 2
Loading