A framework for gardening LLM prompts
Explicitly self-described as rough and early: "⚠️ This library is under active development. There are many rough edges." Package version is 0.0.4 with no tagged releases. The README warns that local profile storage formats "will likely change often, which means you might lose your state." Last commit landed October 2023 — roughly 2.5 years dormant as of this appraisal. 56 open issues with no visible triage activity. This is firmly abandoned alpha-ware.
The README punches above its weight for a 22-star project: covers installation, CLI usage with real output examples, webapp walkthrough, npm package embedding with a code snippet, Polymath library import, seed packet authoring, and remote seed execution. A companion example repo (jkomoros/prompt-garden-example) is referenced. However there is no formal docs site, no API reference beyond the README and VSCode JSON schema hints, and no changelog — breakage is self-acknowledged as undocumented.
TypeScript throughout is a solid choice. Zod is used for schema validation. hnswlib-node provides local HNSW vector search for associative memory. Build pipeline uses tsc + rollup with clean and copy steps. Mocha tests exist (test/base, test/template). One notable hygiene issue: eslint and typescript appear in dependencies rather than devDependencies, inflating the production install footprint. Lit web components are used for the webapp. The declarative seed-graph JSON format is a thoughtful design.
No commits since October 2023 (~2.5 years). 56 open issues — a backlog that was never caught up with. No pull request merge activity visible. The Discord was set up but there is no signal it remains active. For all practical purposes this project is abandoned without being archived, which is the worst-case maintenance state.
22 stars and 3 forks after almost two years on GitHub is negligible. The npm package (prompt-garden) was published but no downstream dependents are visible. The referenced example repo is the only known consumer. The space it targeted — structured LLM prompt chaining — was quickly overtaken by LangChain, LlamaIndex, and DSPy ecosystems that attracted orders-of-magnitude more adoption.
Overall: 1.6/5
Category: LLM Prompt Composition Known alternatives in vault: None — this is the first repo appraised for this category. The nearest neighbor is geeknik--HyperTune (LLM Hyperparameter Optimization) but that addresses a different problem. Differentiation: Prompt Garden's core differentiator is the declarative seed-graph format: prompts are defined as JSON packets with typed, composable seeds that reference each other, enabling version-controlled, shareable, forkable prompt programs. It also includes a local associative memory layer (HNSW vectors via hnswlib-node) baked in rather than bolted on. The webapp editor with JSON schema hints lowers the barrier to authoring. Nothing else in the vault combines prompt composition with local vector memory in a single package. Gap or crowd: Gap — no other repo in the vault covers structured LLM prompt composition. However the category itself was rapidly commoditized between 2023 and 2026 by mature open-source frameworks, making the marginal value of this specific implementation low despite the gap.
Score: 2/5
Harvestable: The seed-packet JSON schema design (typed, composable, declaratively namespaced prompt graphs) is a well-thought-out pattern worth studying. The profile abstraction separating browser vs. node storage is clean. The hnswlib-node integration pattern for local vector memory in a TypeScript project is directly applicable. The Polymath library import pipeline shows a useful "ingest external knowledge corpus → associative memory" workflow.
Integration path: Could be embedded via the npm package as a prompt-execution engine in a PAI tool layer. The CLI binary (garden) could be invoked as a subprocess skill. The seed-packet format could inform how PAI prompt templates are versioned and composed. Direct use of the webapp as a PAI-internal prompt editor is possible but impractical given the stale codebase.
Overlap with existing: No strong overlap with current vault entries. The Personal AI Memory category (garrytan--gbrain, NorthwoodsSentinel--loam) handles retrieval; prompt-garden handles composition. These are complementary rather than redundant.
Adoption cost: Significant — the codebase is 2.5 years stale in a fast-moving domain, carries no release versioning, and the README warns of frequent breaking changes. Any integration would require forking and updating API bindings (OpenAI API has changed substantially since 2023), auditing the dependency tree, and porting the seed-graph executor. Treating it as idea-harvest rather than live dependency is the realistic adoption path.
Prompt Garden is a thoughtful early attempt at a problem that became crowded and then mature between 2023–2025. Its declarative seed-graph concept anticipated ideas that LangGraph and similar frameworks later productionized at scale. The local vector memory integration and the namespace-scoped seed packet format show genuine design care. However, the project stalled at 0.0.4 with no releases, an accumulating issue backlog, and zero commits for 2.5 years — placing it firmly in the "interesting fossil" tier rather than live infrastructure. Suitable for pattern-mining and conceptual reference only; do not take a runtime dependency.