From the field · May 2026
Most prompt libraries die in week 2. Here's what keeps one alive — and what Claude can do to help maintain it.
The problem
The failure mode is predictable. You find a promising prompt on Twitter, or you write one that worked well in a session. You paste it into a Notion page called "Prompts." Two weeks later, you can't remember what it was for, it doesn't quite work without the context you had when you wrote it, and the Notion page has 12 more entries you also can't use without effort. You stop checking it.
The problem isn't the prompts. It's the format. A prompt library that saves only prompt text is almost useless. Prompt text without context notes is like saving a recipe without the serving size or the occasion it's meant for. You can't tell at a glance whether this prompt is for you, right now, with the project you're working on.
The fix is structural: save more than the prompt.
The format
Every prompt I save has four things:
The prompt text itself comes after these. The metadata is what makes it reusable. Without it, you're opening the file and reading the prompt cold — which means reconstructing the context in your head every time, which means you eventually stop bothering.
prompts.md — real saved prompt entry
Patriola's Guide to Claude
## inbox-triage
**What it's for:** Sort 10–30 email subject lines into three categories:
decision needed, needs reply, can ignore.
**Context it needs:** A plain-text list of sender + subject line from
the last 18–24 hours. Optionally: the first line of ambiguous emails.
**What to expect:** Returns three labeled sections with one-line
explanations per item. Tends to over-flag "decision needed" on
invoice follow-ups — correct those and add calibration examples to
the system prompt if it keeps happening.
**Prompt:**
You are my inbox triage assistant. Sort each email I give you into
exactly one of three categories:
DECISION NEEDED — Someone is blocked waiting for my answer, OR I need
to choose between options with real consequences.
NEEDS REPLY — Someone asked me something, but nothing is blocked.
CAN IGNORE — Newsletters, announcements, receipts, confirmations,
automated notifications, anything where no reply is expected.
For each email: [CATEGORY] — Sender: Subject + one sentence why.
Be conservative with DECISION NEEDED.
Storage
Under 20 prompts: a single prompts.md file in your project folder or home directory. Fast to search with Ctrl+F, fast to edit, no app to open. The format above, repeated for each entry.
Over 20 prompts: a prompts/ folder with one file per category — prompts/writing.md, prompts/code-review.md, prompts/research.md. Same per-entry format, just split by theme so the files stay scannable.
What I'd avoid: any dedicated prompt app, browser extension, or database. The friction of opening a separate tool is enough to kill the habit for most people. Keep it where your other work files are.
Maintenance
Every few months, paste your prompt library into Claude and ask it to flag anything that looks unused, outdated, or redundant. Claude will identify prompts with overlapping jobs, entries where the "what to expect" section suggests the prompt wasn't reliable, and anything that references tools or workflows you've moved away from.
You make the final call on what to keep or cut — Claude just surfaces candidates. This takes about five minutes and keeps the library from becoming a graveyard.
Adding a new prompt with Claude's help
I want to add a new prompt to my library. It's for reviewing a piece of writing and identifying where the argument loses momentum — places where the reader might disengage. Can you help me write the full entry in the format I use?
My format: Name, What it's for, Context it needs, What to expect, then the prompt.
Name: argument-momentum-review
What it's for: Identify sections where an argument loses forward motion — where the reader is likely to disengage or skim.
Context it needs: The full draft text. Works best on pieces of 400–2000 words; longer pieces should be submitted section by section.
What to expect: Returns a list of specific passages (quoted) with a one-sentence diagnosis for each. Tends to flag transitional paragraphs that could be tightened and sections where abstract claims run longer than 2–3 sentences without a concrete example. Rarely misses genuine low-momentum sections, but occasionally flags tight prose as "slow" when it's actually doing structural work — read the diagnosis before cutting.
Prompt:
Review the following writing for momentum. Identify every section where the argument loses forward motion — where a reader might disengage or start skimming. For each section: quote the passage, diagnose why momentum drops, and suggest one specific fix. Do not rewrite the passage unless I ask. Focus on structural momentum (argument progression) not surface prose.
"A prompt worth saving has been tested, has a specific job, and would be genuinely hard to re-write from scratch. If it takes 30 seconds to reconstruct, don't bother saving it."
Next step
Book 1 ships with a starter prompts.md file — 12 tested prompts in the format above, covering inbox triage, task prioritization, writing review, end-of-day close, and session loading. You can use them directly or treat them as calibration examples for building your own.
Newsletter
No schedule. Just the things worth writing down — session memory, hooks, automations, multi-agent setups.