Loading...
Loading...
Found 56 Skills
Local proxy that lets OpenAI Codex CLI/desktop talk to MiMo, DeepSeek, and other LLMs via Responses API translation
Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.
Provides image recognition capabilities for non-multimodal models (such as pure text models like deepseek-v4-pro, GLM-5.1, mimo-v2.5-pro, etc.). This skill is automatically triggered when the main model cannot recognize images, when users send screenshots/design drafts/UI screenshots for analysis, or when users say 'Look at this image', 'Analyze this screenshot', 'What's wrong with this image'. It also applies to any scenario where users paste images but the current model does not support image input. Supports simultaneous recognition of multiple images, with primary-backup fallback achieved by configuring multiple image recognition models. It can also be manually triggered using the commands /skill:vision-support or /vision. Iron Rule: The models configured for this skill are only used for image content recognition and will never participate in main logical reasoning. Note: If the current model is itself a multimodal model (such as Claude Sonnet 4, GPT-4o, Gemini, etc. that can directly recognize images), do not use this skill; let the main model recognize directly.
Run a complete repository-native, spec-anchored development loop inspired by DeepSeek Harness. Use for feature development, bug fixes, simplification, architecture, process, or testing work that changes behavior, contracts, structure, test strategy, durable formats, or decision rationale; when asked to write, review, continue, implement, reject, supersede, archive, or verify a spec/Agent Note/RFC/ADR; or when code, tests, current-state docs, and decision rationale must converge in one change. Also use for read-only spec-to-code drift checks. Do not create a spec for a purely mechanical local edit that changes none of those things.
Use when reviewing a pull request in the deepseek-harness repo — orients the reviewer to this codebase's standards (AGENTS.md conventions, defensive patterns, ADRs, quality gates) and the review-specific checks that code alone can't show
Parallel codebase onboarding — distilled understand-anything pipeline (scan + git history + batch analyze + layers/tour, no plugin), then domain enrichment (Claude), wiki + HTML via opencode+DeepSeek Flash v4.
Use when writing, reviewing, restoring, trimming, or auditing prose in the deepseek-harness repo, including deciding where documentation or comments are required across Markdown, JSDoc, code and test comments, prompts, descriptions, diagnostics, and CLI or UI strings.
Internal guidance for composing prompts that Pi runs (DeepSeek by default) handle reliably for coding, review, diagnosis, and research tasks