AI Prompt Manager and Local Prompt Library

PhraseVault® is a local AI prompt manager that stores your reusable prompts in a searchable library on Windows, macOS and Linux, with no account required. On Linux, PhraseVault is an x86_64 build for Omarchy and Arch with Hyprland. Automatic paste and the summon shortcut need Hyprland, and the AppImage is best-effort. See Install PhraseVault on Linux. Press Ctrl+. on Windows or Cmd+. on a Mac, search for a prompt, and press Enter to paste it into ChatGPT, Claude, Microsoft 365 Copilot, or another text field. PhraseVault manages the text; the AI tool you choose handles the request.

Why chat history is a poor prompt library

You write a prompt that works, get a good result, and then lose the exact wording in the thread. Next week you retype it from memory and spend time adjusting it. Notes and documents keep the text, but you still need to switch windows and copy it. ChatGPT and Claude offer project instructions for recurring context, while Microsoft 365 Copilot lets you save prompts within its own service. A desktop library is useful when you want to reuse the same wording across several tools, including local AI apps.

Insert a saved prompt

Save prompts as Text, so the chat box receives the characters you wrote, including any Markdown marks. Click where the prompt should go, press the shortcut, type a few letters, and press Enter. Fuzzy search finds the prompt however you word the query, from the name or unprotected content, in any order, including partial words.

On a Mac, allow PhraseVault in System Settings → Privacy & Security → Accessibility. Until you do, the prompt stays on the clipboard.

Placeholders turn a fixed prompt into one you can reuse. Copy the source text first if you use the clipboard.

• {{clipboard}} inserts whatever you copied.
• {{input:Topic}} asks for a line before pasting.
• {{select:formal,friendly}} asks you to pick one.
• {{date}} inserts today's date.

The full set, including time and a multiline field, is in the placeholders guide.

Prompts worth keeping

Summarize a thread:

Summarize the conversation below as:

1. Decisions
2. Open questions
3. Actions, with an owner if one was named If something was not said, write "not specified."

{{clipboard}}

Rewrite as an email:

Rewrite the text below as a professional email. Keep the intent, reduce ambiguity, and stay under 180 words. End with one clear next step.

{{clipboard}}

Meeting notes into actions:

From the notes below, extract:

1. Decisions, with the reason if it was stated
2. Actions as a table: Task | Owner | Due
3. Open questions
4. Three to five bullets on what was actually discussed If an owner or a date is missing, write "TBD".

{{clipboard}}

Review a diff:

Review the diff below. Look for:

• Logic errors, empty inputs, and off-by-one mistakes
• Injection, auth gaps, and data that should not be shown
• Needless work, N+1 queries, missing indexes
• Names and error handling that will confuse the next reader

Order findings by severity, then give a short recommendation.

{{clipboard}}

Improve a prompt:

Improve the prompt below for clarity. Keep the original goal and constraints. Return:

• The revised prompt
• Why you changed it
• Two alternative versions

{{clipboard}}

Pull structured JSON from a text:

From the text below, return JSON only, with this shape: {"summary": "", "people": [{"name": "", "role": ""}], "actions": [{"task": "", "owner": ""}], "risks": []} Use null when the text does not say.

{{clipboard}}

Brainstorm with a choice of tone:

I need ideas for: {{input:Topic}} Write them in a {{select:practical,bold}} tone.

Give me eight ideas. For each, one sentence on why it could work, one sentence on the main risk, and an effort of low, medium, or high. Then pick the best two and say why. Stay specific.

A support reply that stays inside the policy:

Draft a support reply using only the policy text below. If the policy does not answer the question, list what is still missing. Do not invent a rule.

{{clipboard}}

A steady editing voice, pasted at the start of a chat:

You are an editing assistant. House style: plain language, short paragraphs, concrete nouns and verbs. Do not add hype, and do not change the facts unless asked. Point out ambiguous spots instead of guessing.

Plan an SEO content brief:

Create a content brief for the keyword "{{input:Keyword}}".

Include the likely search intent, audience, a title under 60 characters, a meta description under 155 characters, an H2/H3 outline, and internal linking ideas. Suggest five to eight FAQ questions. Only describe them as actual search results if you've checked a current source; otherwise label them as suggestions. Focus on the reader's needs.

Analyze data or a report:

Analyze the {{select:spreadsheet data,report,survey results,log output}} below.

Return the five main findings, patterns, data quality issues, limitations, and next steps. Cite numbers from the supplied data, and flag claims that need verification.

{{clipboard}}

Set up a research assistant:

Help me understand {{input:Topic}}. Distinguish established facts, expert consensus, and speculation. Explain disagreements between sources and use concrete examples. Say when you're unsure instead of guessing. Suggest three things to investigate next.

Code-heavy prompts, and prompts aimed at a coding assistant, are collected in the developer guide.

Name prompts so search can find them

Use descriptive names to organize your prompts. When you save a phrase in the editor, its name must be unique. A pattern that stays readable as the library grows is domain.task.variant.version.

task.rewrite.professional-email.v2 and coding.review.pr-diff.v1 are enough. Type rewrite and every rewrite prompt appears, because search matches any part of the name.

You can also give a prompt one or more labels, such as Writing or Code. Click a label under the search box, or type # at the start of the search or after a space, and pick the label. Several selected labels show only prompts that carry every one of them. The list stays flat, with no folders. See labels and filters.

Keep the version that worked

When a prompt gets better, don't throw away the wording that already produced a good result.

1. Rename the working one so the name ends in v1.
2. Save the new wording as v2 when you change the output shape, the constraints, or the tone.
3. Add a model only when the wording really differs, as in v2-claude beside v2.
4. Put one line at the top of the prompt saying what changed. Keep this note brief, since it's pasted along with the prompt.

If you keep editing the text after it pastes, the saved prompt isn't finished. Update the saved one with the change you keep making. If two prompts do the same job, keep one and use {{select:option1,option2}} for the part that differs. A short library of prompts you trust beats a long list of drafts you scroll past.

Use the library from an AI agent

If you turn it on, an AI agent that runs shell commands can read the same prompts. Open Settings → Security → Command line & AI agents and set Access to Read-only. The agent can then run phrasevault search "followup" --json and phrasevault expand name:followup --set Name=Alex. The command is off by default and talks only to PhraseVault on this computer. There is no AI inside PhraseVault. Setup is in the command line guide.

Share the library with a team

Put the database file in a shared folder and the team inserts the same prompts. A change shows up the next time someone searches, and it isn't built for two people editing the same prompt at the same moment. Agree the name pattern first, and let one person own each area, such as support or coding. Setup is in the team sharing guide.

More examples you can try in the browser are on the AI prompt examples page.

Try PhraseVault free for 14 days — all features included, no credit card or account needed.