Writing your own prompts

A prompt is the set of instructions MacParsley gives an AI model along with a transcript: what kind of document to write, which sections to include, and how it should look. You don't need to be a developer to write one. A prompt is just plain English in a text box.

Write your own when none of the 60+ built-in prompts gives you the layout you want, for example a weekly report in your team's format, a client-call note with your own headings, or a lecture summary that always ends with review questions.

Try the easy things first. The five styles (Standard, Concise, Detailed, Executive Brief, Narrative) change length and tone without any writing, and a built-in prompt can be copied and tweaked. See Summaries and mind maps.

Start from a copy

1. Open Settings → Prompts and find the built-in prompt closest to what you want. Built-in prompts can be read in full but not changed.
2. Click Duplicate & Edit. You now have your own copy, named … Copy.
3. Change what you like and click Save.

To start from nothing, click New Prompt at the bottom of the same window. You can also right-click or double-click a prompt in the list to view, edit or duplicate it.

The editor's fields

What MacParsley already tells the model

Every prompt is sent together with a set of standing rules, so you don't have to repeat them. They say, in short:

• use only what's in the transcript, and never invent names, dates, numbers, owners or deadlines;
• leave out details that weren't given instead of writing "N/A" or "TBD";
• the transcript comes from speech recognition, so fix an obvious mistake only when the context makes it clear;
• use speaker names as given, and don't guess who "Speaker 2" is;
• keep specifics (names, numbers, dates, product names) and the reason someone gave for something;
• attach each fact to the right person and event, and list only tasks that are still open;
• ignore adverts and sponsor reads;
• write in the language of the recording, and turn "next Friday" into a real date;
• reply with the finished Markdown document and nothing else.

The chosen style is added after your instructions (for example "keep it short"). So your prompt only needs to say what the document looks like.

Writing the instructions

A good prompt reads like a brief to a capable assistant who hasn't heard the recording.

1. Say what the document is for and who reads it. "A one-page recap of a customer call for the account manager."
2. List the sections, in order, with a line on what goes in each. Headings in Markdown (## Decisions) make a clear layout.
3. Say when a section applies. "Include Risks only if someone raised a risk." Without that, a model fills every section even when there's nothing to say. Letting sections be left out is the single biggest improvement you can make.
4. Be specific about format. "Bullets, one line each, starting with a bold label." Bold labels followed by a short phrase are easy to scan.
5. Say what to do with tasks. "List action items as - [ ] Task (owner, date), only if an owner was named."
6. Ask for what you want, not against what you don't. "Write in plain sentences" works better than a long list of don'ts.
7. Keep it short. Ten clear lines beat a page. Small models on your Mac follow short prompts much better.

What to avoid

• Repeating the standing rules above. It wastes space and can make smaller models confused.
• Layout notes that look like content. If you write (the decision here) in a template, say clearly that it's a note and should be replaced. MacParsley already tells the model this, but plain wording ("Write one sentence about…") is safer.
• Asking for facts the recording won't contain, like "budget" in a chat about lunch. Use the "only if" wording instead.
• Vague words such as "good" and "professional". Describe the result instead.

Choosing a category and Use when

When you pick Automatic in the Summarize window, MacParsley decides in two steps:

1. It reads the start of the recording and picks the best category, using each category's short description.
2. Inside that category it picks the best prompt, using each prompt's Use when sentence.

So for your prompt to be found:

• Put it in the category that fits. A Meetings prompt is only considered if the recording looks like a meeting. You can also create a new category.
• Write Use when as one plain sentence about the recording, not about the output: "A weekly check-in between a manager and one team member", not "Writes a status report".
• Make it different from its neighbours. If two prompts in a category say nearly the same, MacParsley can't tell them apart. Mention what's distinctive: customer, internal, two people, a lecture with slides.
• If the prompt is only for a special case, or you always choose it yourself, turn off Choose automatically when it fits. The built-in Lenses and Rewrite prompts work this way.

The description of a whole category can be changed in Settings → Prompts (the Edit link under the category name; Restore Original undoes your change).

Long recordings

Recordings too long for the model to read at once are split. MacParsley takes notes on each part and writes the document from the notes, so most prompts need nothing extra.

If a prompt's layout would be wrong when written from notes (for instance it contains a timeline), use When combining parts to say how the parts should be joined: "Merge the parts into one document with the same sections. Remove duplicates. Keep the time order." Leave it empty if you're unsure.

Try it and improve it

1. Open a meeting you know well and click Summarize.
2. Choose your prompt (and a style, and a model).
3. Read the result against what you remember. Is anything invented? Anything missing? A section that shouldn't be there?
4. Change one thing in the prompt, then summarize again. MacParsley keeps each summary as a version, so you can compare them and keep the better one.
5. Try it on a different meeting too. A prompt that works on one recording may not on another.

If a result is wrong in the same way every time, the fix is usually one clear sentence: "Don't list a task unless someone agreed to do it."

Remember that the model matters. A prompt that works with a large cloud model may need to be shorter and simpler for a model on your Mac. Test with the one you'll really use.

A worked example

Name: Weekly Team Report Category: Meetings Use when: A weekly team meeting where each person reports progress and blockers. Choose automatically: on Sensitive: off

Instructions:

Write a one-page report of this weekly team meeting for people who weren't there.

## Summary
Two or three sentences: what the meeting was mostly about.

## Progress
One bullet per person who reported. Start with their name in bold, then one line on what they finished.

## Blockers
One bullet per problem raised, with who is affected and what was said would fix it. Leave this section out if nobody raised a blocker.

## Action items
Use "- [ ] Task (owner, due date)". List only tasks that are still open and that someone agreed to do. Leave the section out if there are none.

Notice what it does: says who the reader is, gives each section a purpose, says when a section may be left out, and doesn't repeat the standing rules.

Where prompts are stored

Your prompts are saved as ordinary text files in the Prompts folder of MacParsley's data folder:

~/Library/Application Support/MacParsley/Prompts/

Show Folder in Settings → Prompts opens it. They're shared by every library on this Mac. That means you can:

• Back them up or copy them to another Mac by copying the files;
• Share one with a colleague by sending the file; put it in the same folder and click Reload in Settings → Prompts, or restart MacParsley;
• Edit one in any text editor.

A prompt file is plain text: a short header between two --- lines, then the instructions.

---
id: weekly-team-report
name: Weekly Team Report
category: Meetings
useWhen: A weekly team meeting where each person reports progress and blockers.
autoSelect: true
sensitive: false
general: false
version: 1
---

Write a one-page report of this weekly team meeting…

## When combining parts

(optional text for very long recordings)

• A file needs an id, a name and a category to load; without them it's ignored.
• The id is the prompt's identity. Make it unique, with no spaces. If your file has the same id as a built-in prompt, yours replaces the built-in one everywhere. Delete your file to get the original back.
• autoSelect, sensitive and general are true or false. (general: true marks a catch-all prompt, usable as the fallback.)
• Everything after the second --- is the instructions. A line reading ## When combining parts starts the optional section for long recordings.

Related pages

• Summaries and mind maps: styles, versions and folder defaults
• AI models and privacy: what the Sensitive switch protects
• Settings reference

← The Summaries tabWhich should I use? →