Push a model like GPT or Gemini to produce higher quality code, design, or writing output.
Run a full evidence based review pipeline on a deliverable before shipping it.
Use quick mode to get a single-response quality lift inside Claude Code or another chat tool.
Apply written craft standards to keep AI generated marketing or product copy from sounding generic.
| apoorvjain25/frontier | joeseesun/qiaomu-userscripts | openclaw/libopus-wasm | |
|---|---|---|---|
| Stars | 77 | 77 | 78 |
| Language | JavaScript | JavaScript | JavaScript |
| Setup difficulty | easy | easy | easy |
| Complexity | 3/5 | 2/5 | 2/5 |
| Audience | developer | general | developer |
Figures from each repo's GitHub metadata at analysis time.
Runs as a Claude Code skill and plugin, or as a plain prompt file in other chat surfaces like Cursor or Windsurf.
frontier is a Claude Code skill and plugin that tries to make other AI models, like Opus, Sonnet, GPT, or Gemini, produce work closer to what the strongest frontier model would ship. The idea is that a model asked to just "make it great" tends to fall back on its training average, the same generic layouts and hollow marketing phrases. frontier's answer is not more intelligence but more method: it defines a quality standard before generating, produces several independent attempts instead of polishing one, checks the result against real evidence, and refuses to stop at "looks done". At its core are 21 written craft standards covering areas like design, code, writing, research, product, security, marketing, and more. Each standard defines what excellent looks like in checkable terms, with a list of overused AI phrasing to avoid and a verification checklist specific to that domain. The tool is invoked with a single command, frontier followed by the deliverable and a mode: quick, full, or gate. In quick mode, a single response is enough. The model reads the matching craft standard, drafts against it, then runs one fresh eyes review of its own output and fixes what it finds, all in one reply. Full mode runs a longer five phase pipeline: scoping and picking the right standard, generating three to five distinct candidate attempts and ranking them, producing the work section by section while re-reading the rubric, gathering evidence like screenshots or test runs, then running repeated fresh eyes judging passes until two clean passes happen in a row, followed by a final taste gate review. Findings from these judging passes are reported in a fixed, parseable format tied to a specific location and rubric line, rather than vague feedback. The project's own README documents that frontier was tested on itself before release and did not pass on the first attempt, with the failed passes and fixes published as examples in the repository. It works not only inside Claude Code but also in other chat surfaces like Cursor, Windsurf, and aider using a plain prompt file. The project is released under the MIT license.
A Claude Code skill with 21 written quality standards that pushes AI models like GPT or Gemini to produce more polished, frontier-grade work.
Mainly JavaScript. The stack also includes JavaScript, Claude Code, Markdown.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.