Track 01 · Foundations · Lesson 04

SAMPLE LESSON

Build a Reusable Prompt Library

What you'll be able to do after this

✓Set up a versioned prompt library folder in under 10 minutes using a naming system you can search in 5 seconds
✓Immediately copy-paste any of 5 battle-tested prompt templates for the jobs you run every week
✓Know exactly which model to route each job to — Gemini vs Claude — without guessing
✓Run a weekly 5-minute library review that makes your prompts compound in quality over time

The Backbone of Your AI OS

Most people use AI the way they use a search engine: type something, get something back, close the tab, forget it. Every task starts from scratch. Every prompt is disposable. The output is inconsistent because the input never compounds.

A prompt library is the opposite habit. It's a folder of tested, version-controlled prompts — one file per job, named so you can find any prompt in 5 seconds. The first time you write a prompt, you invest 10 minutes. Every use after that is 10 seconds. After 3 months, the library is worth more than access to any single AI tool.

This is the backbone of a personal AI OS because everything else depends on it. Workflows are prompts chained together. Agents are prompts given memory and tools. If your prompts are scattered, your OS is unstable. If they're versioned and maintained, every new model release makes you more capable — not confused.

“The library is not the tool. It's the expertise that survives tool changes.”

Folder Structure and Naming

The structure matters less than the habit of having one. Use this exact layout to start — it works for 5 prompts and scales to 50 before you need subfolders.

YOUR LIBRARY STRUCTURE

prompts/
├── summarize.md          # Document or article → key points
├── rewrite-edit.md       # Draft → polished, tight version
├── extract-actions.md    # Meeting transcript → action table
├── research-brief.md     # Sources → structured research brief
├── decision-helper.md    # Decision context → options + recommendation
└── _ARCHIVE/             # Prompts you've retired (keep, don't delete)

Naming rules: verb-noun, kebab-case, no dates in the filename. Dates belong inside the file as a comment when you update it — not in the name, where they go stale and create noise.

Each file follows the same internal structure:

INSIDE EACH PROMPT FILE (example: summarize.md)

# summarize.md
# updated 2026-07 — added confidence flags for findings
# route: long docs → Gemini | short docs → Claude

[prompt text here]

One comment line per update. That's your changelog. Future you will read it in 3 seconds and understand exactly why the prompt changed.

Your Starter Library: 5 Copy-Paste Templates

These are fully specified prompts for five jobs almost every knowledge worker runs weekly. Copy them as-is into the matching file. Variables in{curly braces} are yours to fill before each run — everything else should stay fixed until you have a reason to change it.

Template 1 — Summarize

ROUTING RULELong documents (10k+ words) → Gemini · Short documents → Claude Sonnet 5

COPY → summarize.md

You are a senior analyst summarizing {document_type} for {audience}.

Read the document below and output:

## Core Idea (2 sentences)
The single most important thing to understand. No hedging.

## Key Points (max 7, numbered)
One sentence each. Concrete, specific, zero padding.
If a point requires "it depends" — say what it depends on.

## What This Changes
1–2 bullets: what should the reader do differently, or think about differently, after reading this?

## Open Questions
Gaps, unsupported claims, or things worth verifying before acting on this.

Length of this summary: {length_constraint — e.g. "fit on one page" or "under 300 words"}

Document:
{paste_document}

Template 2 — Rewrite / Edit

ROUTING RULEClaude Sonnet 5 always — preserves voice better than any other model

COPY → rewrite-edit.md

You are a professional editor. Your job is to improve this draft without changing the author's voice or core argument.

Author's voice: {describe in 2–3 words, e.g. "direct, technical, no hype"}
Audience: {who reads this}
Goal of this piece: {what the reader should do or think after reading}

Edit for:
1. CLARITY — cut sentences over 25 words in half; replace jargon with plain language
2. PUNCH — delete the first sentence if it's a throat-clear; delete the last sentence if it restates what was just said
3. SPECIFICITY — replace vague words ("many", "important", "significant") with concrete ones; if you can't be specific, flag it

Output:
- The edited version in full
- A 3-bullet list: the most significant changes you made and why

Do not invent facts, change the author's position, or add examples they didn't provide.

Draft:
{paste_draft}

Template 3 — Extract Action Items

ROUTING RULEClaude Sonnet 5 — structured table output is more reliable than Gemini for this job

COPY → extract-actions.md

You are extracting structured action items from a meeting transcript. Be ruthless: only log what was explicitly agreed.

Output a table:
| Owner | Action | Due Date | Depends On |
Each row must be specific enough to execute without re-reading the notes.

Rules:
- Owner = a person's name, not a team name
- Action = one clear verb + deliverable ("Send final proposal to Legal" not "follow up")
- If no due date was mentioned: write "UNDEFINED — owner must confirm"
- If ownership is unclear: write "[ASSIGN]" in the Owner column

After the table:

## Open Questions (discussed but not resolved)
Numbered list. Flag anything that's blocking an action.

## Decisions Made
Numbered list. What was agreed. One sentence each.

Transcript:
{paste_transcript_or_notes}

Template 4 — Research Brief

ROUTING RULEGemini 3.x for large source intake (long context) → Claude Sonnet 5 to polish final brief

COPY → research-brief.md

You are a senior research analyst. Read the sources below and produce a structured brief on {topic} for {audience}.

## Executive Summary (100–150 words)
The most important thing a decision-maker needs to know, in plain language. No hedging.

## Key Findings (max 8, numbered)
Format per finding: Claim · Evidence · Source · Confidence (HIGH / MEDIUM / LOW)
Flag every LOW confidence finding with a note on why.

## Implications for {decision_or_question}
3 bullets: what this means for the specific question at hand.

## Gaps and Caveats
What's missing, uncertain, or would change the conclusions if new data came in?

## Sources Drawn On
List each source you cited.

Rules: cite a source for every claim. If you can't find evidence for something, say "no evidence found" — do not fill gaps with plausible-sounding inference.

Sources:
{paste_documents_articles_or_urls}

Template 5 — Decision Helper

ROUTING RULEClaude Opus 5 for complex decisions · Claude Sonnet 5 for routine ones — Gemini is not the right model here

COPY → decision-helper.md

You are a sharp, unbiased decision advisor. Your job is to help me think clearly — not to validate the answer I'm leaning toward.

Decision I need to make:
{describe the decision in one sentence}

Context:
{relevant background — constraints, stakes, timeline, who else is affected}

Criteria I care about (in rough order of importance):
{list 3–5 criteria, e.g. "speed, reversibility, cost, team buy-in"}

Output:
1. OPTIONS — list the realistic options, including any I may not have named
2. TRADE-OFFS — for each option: what it gains, what it costs, what risk it carries
3. KEY UNKNOWNS — what information would most change this decision if I had it?
4. RECOMMENDATION — given my criteria, which option do you recommend and why?
5. BLIND SPOTS — what am I probably not thinking about? Where might I be rationalizing?

Be direct. If one option is clearly better given my criteria, say so. If my framing seems off, challenge it.

Routing Rules at a Glance

The routing note on each template tells you which model to use. Here it is consolidated. Bookmark this table — it's the decision you make before every run.

Job	Default Model	Why
Summarize (short doc)	Claude Sonnet 5	Cleaner structure, better at following exact output format
Summarize (long doc, 10k+)	Gemini 3.x	Long context window; don't chunk what you can read whole
Rewrite / Edit	Claude Sonnet 5	Preserves voice and argument better; Gemini over-rewrites
Extract Action Items	Claude Sonnet 5	Table output is more consistent; less hallucinated owners
Research Brief (intake)	Gemini 3.x	Feed large source sets in one pass; no need to chunk
Research Brief (polish)	Claude Sonnet 5	Tighter prose, better citation discipline in final output
Decision Helper	Claude Opus 5	Reasoning-heavy; nuance matters more than context length

These defaults hold for most cases. The one exception: if a job requires pulling context from 50+ pages of source material, always start with Gemini regardless of the job type. Route to Claude only when the source fits in a normal context window.

The Weekly Review Habit

The library only compounds if you maintain it. The maintenance is 5 minutes, once a week. Do it on Monday morning before you open your inbox — not Sunday, when you won't.

Step 1

Scan for prompts that underperformed

Any prompt you ran this week that produced mediocre output: open it and rewrite the format section. Nine times in ten, that's where the problem is — not the role or task.

Step 2

Capture one new prompt

Did you write a one-off prompt this week that produced great output? Save it now, before you close the tab. Recovery rate for ad-hoc prompts not saved in the moment: approximately zero.

Step 3

Update the routing note

Did you find a better model for a job this week? Update the routing comment at the top of the file. The routing note is part of the prompt — it's wrong until you update it.

Step 4

Archive, don't delete

If you haven't used a prompt in 30 days, move it to _ARCHIVE/. Don't delete it — you might want to understand why it stopped being useful later.

A library of 20 well-maintained prompts — built over 3 months at this pace — is worth more to your productivity than any new AI tool release. Not because the prompts are magic. Because you know exactly what they do, you've tested them, and they're already in front of you when you need them.

Continue Learning

Put these prompts into a real workflow

The next lesson wires these prompts into an end-to-end inbox triage workflow — 20 minutes to set up, 5 minutes every morning to run.

Next: Inbox Triage Workflow →Subscribe for ongoing lessons — $19/mo→