AI prompt templates in a notes tree

You finally get a prompt that works. The role is clear, the constraints stop the fluff, the output format is usable. Two weeks later you need it again — and it is buried in a chat you cannot search by project. Accounts rotate. Threads get summarized into mush. The wording that took an hour to refine is gone.

Three builds ship from this site. SEO Note 2.1 is the current one: an Electron app (about 121 MB installer) for Windows 10/11 64-bit and Linux x64, with Markdown, the tree editor, AI talking to a model you connect — a local one or any OpenAI-compatible endpoint, a web clipper and git sync of the file. For Android 6.0 and newer there is a release-signed APK (~10 MB) of the same interface; on the phone notes live as a folder of Markdown files on the device, and passwords, sync, the clipper and AI answer there with an explicit refusal. The classic build 1.3.34 (~2.5 MB, Win32) stays on the download page for Windows 7, 8 and 32-bit PCs, where 2.1 does not install; its notes file is .seo. All three work offline with no account: pick your build, or read how 2.1 and 1.3.34 differ.

Good prompts are assets. Treat them like you treat checklists and meeting templates: named, versioned, and sitting next to the work they support. SEO Note is a free local outliner for Windows. It ships no model of its own: the AI in 2.1 calls an endpoint you connect — Ollama or LM Studio on this PC, OpenAI, DeepSeek, OpenRouter or your own — and can be switched off. It holds a prompt library you own — one file you open offline, copy to USB, paste into ChatGPT, Claude, or any other chat, and hand to the app's own AI when you need it. No mandatory cloud account, no claim that the notebook “remembers” your workspace for you.

Where the work happens

In SEO Note

• Stable prompt titles
• Roles, constraints, placeholders
• Versions you can compare
• Library next to real projects

In the AI chat

• One-off experiments
• Disposable brainstorms
• Paste + get an answer
• Close when done

Why a tree beats a chat archive

Chat history is chronological. Your brain is topical. When you need “the strict email rewrite for clients,” you want a node with that name — not a scroll through February. A tree also lets you keep siblings: a short version, a long version, a “no humor” variant. That is versioning by hand, which sounds old-fashioned until you realize most “prompt managers” are either another SaaS login or a folder of unmarked text files.

• Stable titles: “Rewrite for clients”, “Debug PHP”, “Exam drill from my bullets”
• Hierarchy: Work / Study / Personal instead of one endless thread
• Search inside one file you can back up like any document
• Privacy by default: nothing leaves the disk until you paste it

A practical library structure

Start smaller than you think. Empty folders feel productive and stay empty. Grow branches when you have two real prompts that belong together.

AI prompts
  Roles
    Editor (clear, short)
    Critic (find gaps)
  Work
    Email drafts
    Meeting summary
    Spec from notes
  Study
    Explain simply
    Quiz me from bullets
  Meta
    Improve this prompt
  Archive
    2026-q1 experiments

Keep Roles thin — reusable personas you drop into other prompts. Keep Meta for the one that rewrites a weak prompt into a better one. Dump failed experiments into Archive with a date so the live tree stays short.

	Chat history only	SEO Note library
Find by project	Hard	Parent node
Keep variants	Easy to lose	Sibling notes
Offline access	Depends on vendor	Local file
What leaves the PC	Whole thread often	Only what you paste

What to put inside each prompt note

A title alone is not enough. Open a note six months later and you should be able to paste it without reconstructing intent from memory.

1. Goal in one line — what success looks like
2. Role and constraints (length, tone, “no fluff”, “do not invent numbers”)
3. Input placeholders like {{paste notes}} or {{code}} so you know what to swap
4. Desired output — bullets, table, checklist, “three options max”
5. Example of a good answer (optional, but it trains your future self as much as the model)
6. Last used date and a one-word verdict: keep / tweak / retire

Rule: name prompts with a verb + domain (“Summarize research paper”), not “New prompt 3”. Untitled nodes are how libraries die.

Workflow that actually sticks

Capture on the day a prompt works. Waiting until “I have time to organize” is how winners vanish into scrollback.

From lucky chat to library node

1 Copy the winning wording the same day

2 Drop it under Inbox / Prompts

3 Rename: verb + domain

4 Move under Work / Study / Roles

5 On improve: keep old sibling a week, then Archive

Once a month, spend fifteen minutes deleting near-duplicates. If two prompts differ only by a soft adjective, keep the stricter one. Soft prompts produce soft answers.

Pairing prompts with real notes

The library is stronger when it sits next to facts. A “Meeting summary” template should live near the project it serves, or at least link to the parent you usually paste from. When you brief a model, paste curated context from your tree — not the whole vault. That habit is covered more fully in local notes as the source of truth for AI.

Students can keep a thin Study branch of quiz and “explain simply” templates beside lecture trees; see study with AI and local notes. For the wider “second brain” picture — structure offline, models on demand — use second brain without the cloud.

When not to bother

• You only ever ask one-off questions and never reuse wording
• Your team already maintains a shared prompt wiki you must use
• You need your notes injected automatically into someone else’s chat UI — SEO Note doesn’t do that; its AI runs in its own windows

Honest limits

• No hand-off button into a chat site, and no embeddings over your vault — 2.1 calls the model endpoint you connect from its own AI buttons, and the MCP server lets an agent read and edit the notes while the app runs
• You paste context yourself — that is the privacy feature, not a bug
• Huge pasted dumps belong in Archive; keep the live tree for reusable templates
• Desktop Windows focus: AI answers with an explicit refusal in the 2.1 Android build, so phone-first prompt packs need a different tool

Mini FAQ

Should every prompt be rich text? Plain text is usually enough. Use formatting only if you need a checklist you tick while editing.

One file or several? One file with bookmark panels (Work vs Personal) beats five orphaned libraries you forget to open.

What about secrets in prompts? Never bake API keys, passwords, or client names into a reusable template. Put placeholders. Paste sensitive values only when required — and prefer not to.

Where do I get prompt templates for free without a subscription? The template page on this site ships ready-made outlines you import into your own tree. You keep the file, edit any role or constraint, and no account or paywall gates the download.

How do I make one prompt work across different models? Keep the role, constraints, and output format model-agnostic. Avoid vendor-specific syntax (magic tokens, system-role quirks). If a prompt breaks in one model, swap only the formatting wrapper while the core instructions stay in the same node.

Next steps

1. Create an AI prompts parent and an Inbox child
2. Migrate three prompts you already use this week
3. Add placeholders and a desired-output line to each
4. Delete one duplicate you know is weaker
5. Back up the file when the library starts to matter

If you want a quiet place for prompts next to projects and facts, download SEO Note free (no account) and treat the library like any other local file.

Keep going

• Tree templates and PDF guide — role and method skeletons you can copy straight into your own tree.
• If this guide helped, you can say hello or grab ready-made trees on the templates page.

Adjacent tools (VPS for local models, courses, VPN) are not stuffed into every paragraph. If an affiliate link appears, it will be labeled and is never required to use SEO Note.

Related: Notes as source of truth for AI · Second brain offline · Study notes + AI · Tabs to notes tree · FAQ