Teaching Claude to stop writing Keras when you asked for Synalinks
Eleven Claude Skills that keep an LLM coding assistant from hallucinating the wrong framework syntax into your neuro-symbolic project.

What it does
This repo ships 11 “Claude Skills” — structured prompt bundles that slot into Claude Code, Claude Desktop, or the web UI. Each skill covers one slice of the Synalinks framework (core data models, program graphs, agents, training loops, optimizers, etc.). Claude auto-activates the relevant skill based on what you’re asking about, so it supposedly stops generating plausible-looking but broken mashups of Keras, LangChain, and DSPy syntax.
The interesting bit
The real value isn’t the prompts themselves; it’s the recognition that framework-specific LLM assistance decays fast without guardrails. Synalinks is deliberately Keras-inspired, which turns out to be a trap — the surface similarity makes LLMs confident and wrong. These skills are essentially a syntax firewall with 11 narrow gates.
Key highlights
- 11 scoped skills from
synalinks-core(JSON operators, DataModel basics) tosynalinks-datasets(built-in benchmarks, visualization helpers) synalinks-control-flowcovers unusual neuro-symbolic primitives: XOR guards, branch merging, self-consistency votingsynalinks-optimizersincludes quality-diversity methods (OMEGA, Dominated Novelty Search) — not your usual Adam/SGD territorysynalinks-knowledgehas DuckDB-backed RAG/KAG with hybrid search and entity/relation graphs- Each skill bundles
SKILL.md(scannable guide),references/(deep docs), andscripts/(runnable examples) - Install via copy, symlink, or
.skillzip archive depending on your Claude flavor
Caveats
- The README notes Claude “auto-activates” skills based on
SKILL.mdfrontmatter, but doesn’t explain how reliably this works in practice or what happens with ambiguous queries spanning multiple skills - No benchmarks or user feedback on whether the skills actually reduce hallucination rates — the claim is plausible but unverified in the sources
Verdict
Worth a look if you’re actively building with Synalinks and tired of debugging LLM-generated framework soup. Skip it if you aren’t using Synalinks; the skills are tightly coupled to its specific APIs and won’t transfer.