Tested skills and building blocks for AI agents
A shelf of 20 tested skills for AI agents, split into 158 small building blocks ("primitives"): gotcha fixes, calibrations, fail-safes and tool sequences, each with when to use it, what to do, what you should see and how it fails.
Whole skills — 10 credits each (≈ €1.00)
Pick a skill on the shelf below and get all its primitives as one package.
Primitives — 1 credit each (≈ €0.10)
Open any skill below and pick only the building blocks you need.
Upgrade your own skill
Paste your skill; we pick the 5 primitives from the shelf that fit it best, as one bundle for 5 credits (≈ €0.50).
Upgrade my skillThe shelf
MLX lm fuse and GGUF export
Fuse MLX LoRA adapters into base model and export to GGUF for Ollama import; covers --de-quantize gotcha and chat-template matching requirement
Parameterize DB queries
Always pass values to a query engine (SQL, Cypher, etc.) as bound PARAMETERS, never by string- interpolating them into the query text. String interpolation breaks on embedded quote…
OPRO LLM as optimizer
OPRO: given (prompt, score) history, ask local Ollama to iteratively propose better prompts; pre-ILP draft-generation role in the harvest loop
Bge-m3 triple vector qdrant
Use all three bge-m3 output heads (dense+sparse+colbert) in Qdrant named-vector collection with RRF fusion for maximum local hybrid recall
Build RAG reranker pipeline
Add cross-encoder reranker as second stage after vector retrieval (pgvector/Neo4j-vector) to reorder top-K by true semantic relevance — local, no GPU, no API, using sentence-transf…
Legacy refactor chunk checkpoint
Safe AI-assisted refactor protocol: map file into chunks, get plan approval per chunk, commit+test after each chunk, rollback to last green checkpoint on failure
Agentic design pattern selector
decision matrix mapping task/context type to the right combination of patterns from the 20-pattern agentic design taxonomy; most real systems combine 3-5 patterns; includes per-pat…
MLX VLM local vision inference
Native Apple-Silicon VLM inference via MLX-VLM: zero-server, vision feature caching for 11x batch speedup, CLI + Python API + FastAPI server + LoRA fine-tuning
Ollama structured output pipeline step
Drop-in upgrade for Ollama pipeline steps: pass Pydantic schema as format= to engage XGrammar constrained decoding — 6.4x faster, ~100% parse success, no brace-scanning
Textgrad prompt optimizer
Optimize a prompt (or system prompt) using TextGrad — "textual gradients" that backpropagate natural-language feedback through a pipeline of LLM calls. Wraps the prompt in a tg.Var…
Context pipeline assembly
6-stage context assembly pipeline (Collect→Filter→Compress→Arrange→Format→Assemble) that curates what enters an LLM context window before every model call; the critical missing Arr…
DSPy GEPA reflective optimizer
GEPA (ICLR 2026) reflective two-model prompt optimizer: local inference LM + subscription reflection LM reads failure text to propose targeted instruction rewrites
Bottomup topdown elicitation
Elicit a system's requirements both top-down (general) and bottom-up (concrete instances as probes), up/down-chunking, draining when K probes add zero
DSPy system prompt compiler
DSPy MIPROv2 adapter (Rivest pattern): compile a single LPPromptV2 Neo4j node's prompt_text against a rubric metric; emit winning instruction as ILP finding; per-node prompt-chain …
Git worktree for parallel agents
Create isolated git worktrees for parallel Claude Code / agent windows — git worktree add/remove + auto-dependency-install + cleanup — so multiple agents never conflict on the same…
Local VLM image pipeline QA
QA gate for LP image batches: moondream caption + qwen2.5vl:7b photo-vs-illustration verdict over 27 generated images; enforces default-photographic policy; runs between image driv…
Dia dialogue TTS
Generate natural two-speaker dialogue audio from [S1]/[S2]-tagged transcript using Dia-1.6B (Apache 2.0), with non-verbal cues (laughter/sighs) in one local model pass — drop-in up…
GBNF lazy grammar CoT constrained
GBNF lazy-activation grammar in llama.cpp: model reasons freely in CoT until trigger token, then JSON schema enforced — reason-then-commit pattern for evidence assessment
LightRAG local ollama graphrag
Turnkey local GraphRAG: doc corpus -> KG -> multi-mode retrieval, no paid API.
Instructor pydantic extractor
Wrap any local Ollama or cloud LLM in Instructor to extract Pydantic-schema-conformant JSON with context-aware reask on ValidationError — 95%+ first-retry self-correction