dkackman · Aug 30, 2026 · Aug 30, 2026 · Aug 30, 2026 · Aug 30, 2026 · Aug 30, 2026
diff --git a/.github/copilot-instructions.md b/.github/copilot-instructions.md
@@ -40,6 +40,10 @@ The REPL uses a **persistent worker subprocess** for workflow execution to maint
 
 **Result References:** Use `"previous_result:step_name"` to pipe outputs between steps. System automatically generates all combinations when multiple results exist.
 
+**Constant References:** Use `"constant:module.NAME"` to read a value declared in Python instead of copying it into JSON.
+
+**Prompt References:** Use `"prompt:name"` or `"prompt:folder/name"` to load a stored prompt's text from the `prompts/` library at run time (see `dw/prompts.py`).
+
 **Pipeline Configuration:**
 ```json
 {

diff --git a/.vscode/launch.json b/.vscode/launch.json
@@ -3,10 +3,10 @@
   // Hover to view descriptions of existing attributes.
   // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
   //
-  // The examples tree has subfolders (flux/, tasks/, archive/, projects/), so
+  // The workflows tree has subfolders (flux/, ltx2/, minimax/, tasks/, archive/), so
   // the "current workflow file" configurations below are the ones to reach for:
   // open the JSON and run it, wherever it lives. The prompting configurations
-  // take a path relative to examples/, subfolders included - "flux/FluxDev".
+  // take a path relative to workflows/, subfolders included - "flux/FluxDev".
   "version": "0.2.0",
   "configurations": [
     {
@@ -71,7 +71,7 @@
       "cwd": "${workspaceFolder}/",
       "justMyCode": true,
       "args": [
-        "./examples/${input:workflowName}.json"
+        "./workflows/${input:workflowName}.json"
       ]
     },
     {
@@ -105,15 +105,15 @@
       "args": [
         "--log_level",
         "INFO",
-        "./examples/${input:workflowName}.json"
+        "./workflows/${input:workflowName}.json"
       ]
     }
   ],
   "inputs": [
     {
       "id": "workflowName",
       "type": "promptString",
-      "description": "Workflow path relative to examples/, without .json (e.g. ZImage or flux/FluxDev):",
+      "description": "Workflow path relative to workflows/, without .json (e.g. ZImage or flux/FluxDev):",
       "default": "??"
     }
   ]

diff --git a/CLAUDE.md b/CLAUDE.md
@@ -13,11 +13,11 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 bash ./install.sh && source ./activate
 
 # Run a workflow
-python -m dw.run examples/ZImage.json
-python -m dw.run examples/ZImage.json prompt="a cat" num_images_per_prompt=4
+python -m dw.run workflows/ZImage.json
+python -m dw.run workflows/ZImage.json prompt="a cat" num_images_per_prompt=4
 
 # Validate a workflow against schema
-python -m dw.validate examples/ZImage.json
+python -m dw.validate workflows/ZImage.json
 
 # Basic system test (torch, diffusers import check)
 python -m dw.test
@@ -28,7 +28,7 @@ python -m dw.repl
 # HTTP server + web UI (http://127.0.0.1:8765, API docs at /docs)
 python -m dw.serve
 
-# Run all tests (580+ tests)
+# Run all tests (1,600+ tests)
 pytest -v
 
 # Run a single test file or test
@@ -52,7 +52,10 @@ over SSE, and persists history to `~/.diffusers_helper/jobs.sqlite`. The SPA liv
 `ui/` (Svelte 5 + Vite; `npm run build` outputs `ui/dist`, which the server serves).
 Introspection endpoints (`dw/introspection.py`) describe pipeline/class signatures for
 the editor's forms. `dw/hub_cache.py` inventories and deletes from the HF hub cache
-(the UI's Models page). Front-end checks from `ui/`: `npm run check`, `npm run lint`,
+(the UI's Models page). `dw/prompts.py` + `dw/server/enhancers.py` back the
+Prompts page: a stored-prompt library (CRUD under `--prompt-dir`) with an
+enhance-with-AI panel that queues an inline workflow as an ordinary job.
+Front-end checks from `ui/`: `npm run check`, `npm run lint`,
 `npm test`, `npx playwright test` (e2e, starts its own server). See docs/SERVER.md.
 
 Packaging: `pyproject.toml` (console scripts dw-run/dw-validate/dw-repl/dw-serve/dw-test);
@@ -75,6 +78,11 @@ The REPL (`dw/repl.py`) uses a **persistent worker subprocess** (`dw/worker.py`)
 - Values prefixed with `constant:` read a value declared in python rather than copying it
   into JSON: `"constant:diffusers.pipelines.ltx2.utils.DISTILLED_SIGMA_VALUES"`. Resolved
   in `realize_args`, validated by `validate_constant_name()`; anything callable is refused
+- Values prefixed with `prompt:` load a stored prompt's `text` from the prompt library:
+  `"prompt:name"` or `"prompt:folder/name"`. Resolved in `realize_args` (`dw/prompts.py`),
+  rooted at the library rather than the workflow file. The library is `DW_PROMPT_DIR` /
+  `--prompt-dir`, else `./prompts` if it exists, else found by walking up from the
+  workflow file's directory
 
 ### Quantization Support
 
@@ -115,6 +123,7 @@ All entry points use `dw/security.py`. When adding features:
 - **Built-in workflows** need explicit argument mapping: `"prompt": "variable:prompt"`
 - **MPS differences from CUDA**: no autocast, no bitsandbytes, no flash_attn, no triton, no torch.compile. Model offloading has less benefit on unified memory.
 - **`{}`-escaped strings** in JSON arguments: `"{nf4}"` stays as string `"nf4"`, without braces it would try to load as a type
+- **A stored prompt's `text` may not begin with a reference prefix** (`variable:`, `previous_result:`, `constant:`, `prompt:`) — the engine rejects it to prevent double resolution or iteration expansion
 - **Audio+video muxing**: pipelines that generate audio alongside video (LTX-2) have the two muxed into one `video/mp4` file with PyAV in `result.py`
 
 ## JSON Workflow Structure
@@ -146,4 +155,4 @@ Steps can also have `"task"` (with `command` + `arguments`) or `"workflow"` (wit
 
 Pipeline configuration options include: `pre_load_modules` (import modules before loading), `sdnq_optimize` (SDNQ quantized matmul), `enable_attention_slicing`, `disable_attention_slicing`, `attention_backend`, `group_offload`, `enable_layerwise_casting`.
 
-File paths in workflows are relative to the workflow file. Built-in workflows use `"builtin:filename.json"` (resolves to `dw/workflows/`).
+File paths in workflows are relative to the workflow file. Built-in workflows use `"builtin:filename.json"` (resolves to the packaged `dw/workflows/` — distinct from the top-level `workflows/` folder of runnable examples).
diff --git a/README.md b/README.md
@@ -21,6 +21,7 @@ A declarative workflow engine and web UI for the [Hugging Face Diffusers library
 - **Quantization** — BitsAndBytes, TorchAO, GGUF, SDNQ, optimum-quanto
 - **Inference acceleration** — TeaCache, FirstBlockCache, FasterCache, MagCache, TaylorSeerCache
 - **Prompt weighting** — A1111-style `(word:1.5)` syntax with long prompt support
+- **Prompt library** — store prompts once in `prompts/` and reference them from any workflow as `prompt:name` or `prompt:folder/name`, with a web UI for browsing, editing, and AI-enhancing them
 - **LoRA and IP-Adapter** support
 - **Composable workflows** from multiple JSON files with `builtin:` references
 - **Utility tasks** — upscaling, face restoration, segmentation, captioning, frame interpolation, QR codes, and more
@@ -64,6 +65,8 @@ Everything the engine does, in a browser backed by a persistent GPU worker — m
 
 ![The gallery with generated images and videos](docs/img/ui-gallery.jpg)
 
+**A prompt library shared by every workflow.** Store a prompt once, reference it anywhere as `prompt:name` — the Prompts page browses, edits, and filters the library, and an *Enhance with AI* panel expands an idea into a full prompt with a local language model.
+
 **A model manager for the disk your models actually consume.** The Hugging Face hub cache, inventoried: sizes, revisions, last-used dates, free space — download new models by id with live progress, delete with one click.
 
 ![The model manager listing cached models with sizes](docs/img/ui-models.png)
@@ -75,14 +78,14 @@ Jobs queue, stream progress live (per denoising step), cancel cooperatively, and
 ### Run a Workflow
 
 ```bash
-python -m dw.run examples/flux/FluxDev.json
-python -m dw.run examples/flux/FluxDev.json prompt="a cat" num_images_per_prompt=4
+python -m dw.run workflows/flux/FluxDev.json
+python -m dw.run workflows/flux/FluxDev.json prompt="a cat" num_images_per_prompt=4
 ```
 
 ### Validate a Workflow
 
 ```bash
-python -m dw.validate examples/flux/FluxDev.json
+python -m dw.validate workflows/flux/FluxDev.json
 ```
 
 ### Interactive REPL
@@ -242,7 +245,7 @@ a (photorealistic:1.4) portrait with (bright red hair:1.3) and [freckles]
 
 **Interactive schema browser:** [View Schema](https://json-schema.app/view/%23?url=https%3A%2F%2Fraw.githubusercontent.com%2Fdkackman%2Fdiffusers-workflow%2Frefs%2Fheads%2Fmaster%2Fdw%2Fworkflow_schema.json)
 
-See [examples/](examples/) for more workflow files.
+See [workflows/](workflows/) for more workflow files.
 
 ## Documentation
 
@@ -256,6 +259,7 @@ See [examples/](examples/) for more workflow files.
 - [LoRA](docs/LORAS.md) — Loading and stacking LoRA adapters
 - [IP-Adapter](docs/IP_ADAPTER.md) — Image-prompt conditioning
 - [Prompt Weighting](docs/PROMPT_WEIGHTING.md) — A1111-style syntax
+- [Prompt References](docs/WORKFLOW_GUIDE.md#prompt-references) — The stored prompt library and `prompt:` references
 - [Tasks](docs/TASKS.md) — Image processing, ControlNet preprocessors, utilities
 
 ### Reference

diff --git a/docs/ACCELERATION.md b/docs/ACCELERATION.md
@@ -42,7 +42,7 @@ Simplest and broadest support. Compares first-block residuals to decide whether
 
 Higher threshold = more speedup, more quality loss. Start with `0.05` and increase to taste.
 
-**Example:** [FluxDevFirstBlockCache.json](../examples/flux/FluxDevFirstBlockCache.json)
+**Example:** [FluxDevFirstBlockCache.json](../workflows/flux/FluxDevFirstBlockCache.json)
 
 ### MagCache
 
@@ -165,7 +165,7 @@ Some models have multiple variants with different coefficients:
 }
 ```
 
-**Example:** [FluxDevTeaCache.json](../examples/flux/FluxDevTeaCache.json)
+**Example:** [FluxDevTeaCache.json](../workflows/flux/FluxDevTeaCache.json)
 
 ## Cache vs TeaCache
 
@@ -206,7 +206,7 @@ A component can also pin its backend persistently instead, via `set_attention_ba
 
 Prefer the pinned form for a compiled component - the per-call context manager switches implementations under the compiled graph and forces a recompile on every run.
 
-**Example:** [Flux2Dev.json](../examples/flux/Flux2Dev.json), [hunyuan15.json](../examples/archive/hunyuan15.json), [Wan22TI2V5B.json](../examples/archive/Wan22TI2V5B.json)
+**Example:** [Flux2Dev.json](../workflows/flux/Flux2Dev.json), [hunyuan15.json](../workflows/archive/hunyuan15.json), [Wan22TI2V5B.json](../workflows/archive/Wan22TI2V5B.json)
 
 ## Attention Slicing
 
@@ -252,7 +252,7 @@ Typical gains are 1.3-1.5x on diffusion transformers, and compilation stacks wit
 - **Don't combine `fullgraph` with a `cache`**: the cache hooks decide skip-or-compute per step, a data-dependent branch diffusers wraps in `torch.compiler.disable` - it needs the graph break that `fullgraph: true` forbids. Compile with the default (partial) graph mode when a cache is active.
 - **TorchAO quantization needs compile to be fast** - see [QUANTIZATION.md](QUANTIZATION.md#torchao).
 
-**Example:** [FluxDevFast.json](../examples/flux/FluxDevFast.json), [FluxTorchAO.json](../examples/flux/FluxTorchAO.json)
+**Example:** [FluxDevFast.json](../workflows/flux/FluxDevFast.json), [FluxTorchAO.json](../workflows/flux/FluxTorchAO.json)
 
 ## Layerwise Casting
 
@@ -277,7 +277,7 @@ Both `storage_dtype` and `compute_dtype` are required. Applied via the component
 
 `"residency": "on_demand"` on a component is the cheap case of the same trade: the model rests in system memory and is moved to the device whole around each of its own calls. That is a bad deal for anything called once per step, and a good one for a VAE called twice a run - it frees the VAE's VRAM for the denoise loop at the cost of two transfers, where group offloading the same VAE would restream it once per decode tile. See [On-demand components](WORKFLOW_GUIDE.md#on-demand-components).
 
-**Example:** [FluxDev.json](../examples/flux/FluxDev.json) (`"offload": "model"`), [ZImage.json](../examples/ZImage.json) (`"offload": "sequential"`), [MiniMaxH3.json](../examples/MiniMaxH3.json) (`group_offload` per component), [MiniMaxH3Ref2VA.json](../examples/MiniMaxH3Ref2VA.json) (`group_offload` for the transformer, `on_demand` for the VAEs)
+**Example:** [FluxDev.json](../workflows/flux/FluxDev.json) (`"offload": "model"`), [ZImage.json](../workflows/ZImage.json) (`"offload": "sequential"`), [MiniMaxH3.json](../workflows/minimax/MiniMaxH3.json) (`group_offload` per component), [MiniMaxH3Ref2VA.json](../workflows/minimax/MiniMaxH3Ref2VA.json) (`group_offload` for the transformer, `on_demand` for the VAEs)
 
 ## TF32 and cuDNN
 

diff --git a/docs/IP_ADAPTER.md b/docs/IP_ADAPTER.md
@@ -77,7 +77,7 @@ Some base models (e.g. Stable Diffusion 3.5) don't ship a default image encoder,
 }
 ```
 
-See [sd35ip.json](../examples/archive/sd35ip.json) for the full workflow.
+See [sd35ip.json](../workflows/archive/sd35ip.json) for the full workflow.
 
 ## Image Argument
 
@@ -105,6 +105,6 @@ Can also reference a previous step's output:
 
 ## Examples
 
-- [FluxIP.json](../examples/flux/FluxIP.json) — Flux with IP-Adapter for style transfer
-- [sd35ip.json](../examples/archive/sd35ip.json) — SD 3.5 (quantized) with an explicit `image_encoder`/`feature_extractor` pair
-- [ip-adapter.json](../examples/archive/ip-adapter.json) — SD 1.5 and SDXL, each with an `AutoPipelineForText2Image` step
+- [FluxIP.json](../workflows/flux/FluxIP.json) — Flux with IP-Adapter for style transfer
+- [sd35ip.json](../workflows/archive/sd35ip.json) — SD 3.5 (quantized) with an explicit `image_encoder`/`feature_extractor` pair
+- [ip-adapter.json](../workflows/archive/ip-adapter.json) — SD 1.5 and SDXL, each with an `AutoPipelineForText2Image` step
diff --git a/docs/LORAS.md b/docs/LORAS.md
@@ -126,6 +126,6 @@ python -m dw.run workflow.json lora="other-user/other-lora"
 
 ## Examples
 
-- [FluxLora.json](../examples/flux/FluxLora.json) — Flux with realism LoRA and variables
-- [lora.json](../examples/lora.json) — SD 3.5 with yarn art style LoRA
-- [bnb_quant.json](../examples/archive/bnb_quant.json) — Quantized model with LoRA
+- [FluxLora.json](../workflows/flux/FluxLora.json) — Flux with realism LoRA and variables
+- [lora.json](../workflows/lora.json) — SD 3.5 with yarn art style LoRA
+- [bnb_quant.json](../workflows/archive/bnb_quant.json) — Quantized model with LoRA
diff --git a/docs/PROMPT_WEIGHTING.md b/docs/PROMPT_WEIGHTING.md
@@ -71,8 +71,8 @@ Any other pipeline whose class name starts with `Flux` (e.g. `FluxKontextPipelin
 ## Example
 
 ```bash
-python -m dw.run examples/flux/FluxSchnellWeighted.json \
+python -m dw.run workflows/flux/FluxSchnellWeighted.json \
     prompt="a (cinematic:1.5) shot of a (dragon:1.3) breathing [smoke] over a (medieval:0.8) castle"
 ```
 
-See [FluxSchnellWeighted.json](../examples/flux/FluxSchnellWeighted.json).
+See [FluxSchnellWeighted.json](../workflows/flux/FluxSchnellWeighted.json).
diff --git a/docs/QUANTIZATION.md b/docs/QUANTIZATION.md
@@ -55,7 +55,7 @@ For 8-bit:
 "arguments": { "load_in_8bit": true }
 ```
 
-**Example:** [bnb_quant.json](../examples/archive/bnb_quant.json)
+**Example:** [bnb_quant.json](../workflows/archive/bnb_quant.json)
 
 ## TorchAO
 
@@ -78,7 +78,7 @@ Common choices: `Int8WeightOnlyConfig` (any CUDA card), `Int4WeightOnlyConfig` (
 `Float8DynamicActivationFloat8WeightConfig` (fastest, requires compute capability 8.9+ -
 RTX 40-series/Ada or newer).
 
-**Example:** [FluxTorchAO.json](../examples/flux/FluxTorchAO.json)
+**Example:** [FluxTorchAO.json](../workflows/flux/FluxTorchAO.json)
 
 **Pair TorchAO with `torch.compile`.** Int8 weight-only and float8 dynamic-activation
 quant types get their fused-kernel speedups only under compilation - uncompiled they are
@@ -117,7 +117,7 @@ GGUF models load from single files using `from_single_file`:
 }
 ```
 
-**Example:** [FluxGGUF.json](../examples/flux/FluxGGUF.json)
+**Example:** [FluxGGUF.json](../workflows/flux/FluxGGUF.json)
 
 ## SDNQ (SD.Next Quantization)
 
@@ -146,7 +146,7 @@ examples use for their large transformers and text encoders:
 - `quantization_device` / `return_device` — where the quantization pass runs and where the finished component lands; quantizing on `cuda` is much faster than on CPU
 - `use_quantized_matmul` — quantized matmul kernels (CUDA/XPU only)
 
-**Examples:** [LTX2.json](../examples/LTX2.json), [MiniMaxH3.json](../examples/MiniMaxH3.json)
+**Examples:** [LTX2.json](../workflows/ltx2/LTX2.json), [MiniMaxH3.json](../workflows/minimax/MiniMaxH3.json)
 
 ### Pre-quantized models
 
@@ -171,7 +171,7 @@ Pre-quantized SDNQ repos load as complete pipelines. The `sdnq` module must be i
 - `pre_load_modules` — Imports sdnq before pipeline loading (registers quantization method)
 - `sdnq_optimize` — Applies quantized matmul to listed components (CUDA/XPU only, skipped on MPS/CPU)
 
-**Example:** [ZImageSDNQ.json](../examples/ZImageSDNQ.json)
+**Example:** [ZImageSDNQ.json](../workflows/ZImageSDNQ.json)
 
 ## Modular Pipelines