A useful Claude prompt library should preserve the instructions you want to run again without preserving every detail of the conversation that happened to produce them.
Markdown files work well for this because they keep the prompt readable, editable, and portable. You can review the source directly, then copy it into Claude when the task comes back.
Separate the reusable core from the current task
Most prompts contain two kinds of information:
- stable instructions that define the job
- changing context for the current run
Keep the stable part in the library. Add the current document, code sample, audience, or deadline when you use it.
For example, a reusable research prompt can define how to compare sources and report uncertainty. The topic and source material should remain inputs, not permanent lines buried inside the prompt.
Use one file shape consistently
A small template makes prompts easier to inspect:
# Compare research sources
Use when: several sources make overlapping or conflicting claims.
## Inputs
- research question
- source material
- required date range
## Prompt
Compare the sources against the research question. Separate agreement,
conflict, and missing evidence. Cite the supplied source for each material claim.
## Output contract
1. Short answer
2. Evidence table
3. Conflicts and uncertainty
4. Next research step
## Claude-specific note
Attach or paste the source material before running the prompt.
The model-specific note stays small. The task, inputs, and output contract remain useful even if you later run the same workflow elsewhere.
Organize by work, not by model
A top-level claude folder feels convenient at first, but it can trap prompts inside a tool label. Prefer folders such as research, writing, development, and planning.
Add Claude to a description only when the prompt truly depends on a Claude-specific behavior. Otherwise, keep the file model-neutral and let the library reflect what you do.
Record the boundary of the prompt
Reusable prompts become easier to trust when they say what they do not cover. Add a short boundary when a prompt should not make a decision, infer missing facts, or handle sensitive material.
Boundary: summarize only the supplied evidence. Do not fill missing dates,
owners, or conclusions from general knowledge.
That line prevents a future user from treating a narrow helper as a broader authority.
Test the file outside its original conversation
Before saving a prompt as reusable, open a fresh conversation and run it with a second example. If it only works because the old thread supplied hidden context, the file is not ready.
Check whether:
- the required inputs are named
- the output is recognizable
- missing information is handled honestly
- another person could tell when to use it
Keep the library independent
Promptlight reads Markdown prompt files from a local vault folder, indexes them for search, and lets you copy the prompt into the tool where you want to run it. There is no direct Claude integration or hosted sync layer. The practical benefit is that the reusable source remains a file you control instead of becoming dependent on one conversation surface.
Start with three Claude prompts you already repeat. Put each in the same file shape, test it in a fresh conversation, and keep only the version that works without hidden context.
For the storage model, see Markdown Prompt Library and Local-First Prompt Manager. For a broader setup, use Build a Local-First Prompt System.
