forkprobe

As the AI skill ecosystem expands rapidly, the real challenge is no longer whether a skill exists, but which skill is actually right for the current task. Choosing based on descriptions alone is risky—you may only discover a poor fit after running the task. ForkProbe is built for this uncertainty: it turns skill selection into an observable trial run. You can send the same task through a baseline and multiple candidate skills side-by-side, then receive a local HTML report that compares full outputs, elapsed time, token estimates, file previews, and AI reviews, so you can pick the winner based on real results rather than marketing copy.

ForkProbe's core strength goes beyond text comparison. It offers seven working modes covering academic writing and polishing, PPTX generation, scientific figure creation, research reports, image prompt and style direction, runnable web pages, and video production. For artifact tasks, each candidate produces a complete package—previews, source files, QA results, and reviews are all shown in the report, not just abstract outlines. Candidate discovery is flexible: it can auto-scan locally installed skills, query the EverMind Skill Hub, search GitHub, or accept explicit local paths, GitHub URLs, or raw SKILL.md links, with content fingerprint deduplication and scenario-based ranking.

After you choose a winner, ForkProbe saves the local verdict and generates a continuation handoff, allowing your agent to continue the real task along the winning skill's style, structure, or produced files. You can also optionally share your skill choice anonymously, contributing statistical priors for community recommendations while only uploading task type, candidate names, and final choice—never task content or outputs.

ForkProbe is ideal for agent users who need to compare multiple skills or pipelines and deliver finished artifacts such as presentations, scientific figures, research reports, image prompts, websites, or videos. It ships as a native DeepSeek Harness plugin and also works with Claude Code, Codex, and other natural-language agent workflows. Task content stays local by default, external sources only receive sanitized scenario terms, and a local-only mode is available for privacy-sensitive use cases.