Prompt Pack — 40 tested prompts with 83 eval cases
Forty production prompts across extraction, classification, support and coding, each with a fixed output shape, two or more eval cases and a runner that checks them against the API.
Short answer. The pack is a ZIP containing 40 prompts, 83 eval cases and a runner. It is a one-time $49 download, not a subscription, and it is delivered from Dashboard → Billing → Add-ons immediately after payment.
USD. If the file does not download, contact support for a replacement or a refund.
Re-download any time from this workspace entitlement.
On this page · 12 sections
Who it is for
Teams shipping an LLM feature who need a starting point that is testable, not a list of clever one-liners.
The problem it removes
- A prompt that works in the playground fails on real inputs.
- Nobody agrees what a correct output looks like, so nobody can test it.
- Every new feature starts from a blank prompt box.
What you get
- 40 prompts: extraction, classification, support and coding
- 83 eval cases with checkable assertions
- prompts.jsonl — one object per line, easy to load
- run_evals.py — runs every case and prints pass/fail
- A model recommendation and a failure note per prompt
How to use it
1. Download the pack
The ZIP is delivered from the dashboard after purchase. It contains the JSONL, the runner, a README and the licence.
2. Run the evals
Set PLUGSKY_KEY and run run_evals.py. The 83 cases pass on the shipped files; the run log is in the ZIP.
3. Adapt and re-run
Change the prompt and the eval together; the eval is the contract, not an afterthought.
curl -sS https://plugsky.com/api/addons/prompts/pack.zip \
-H "Authorization: Bearer $PLUGSKY_SESSION" \
-o plugsky-prompt-pack.zip
See before you buy
Version 2026-11 · 5 files · 13.5 KB. The ZIP includes the 83/83 run log. One prompt and one eval case, copied from the pack:
Categories: tools, grocery, electronics, apparel, home, other.
Return JSON: {"category":"one of those six","confidence":0-1}.
Input: Cordless drill with two batteries
Expected: {"category":"tools"}
The eval asserts category == "tools". If the model returns "power tools", the case fails and the prompt is fixed — that is the loop the pack sells. Licences: commercial use inside your organisation; no resale of the pack itself.
What is in the pack
The pack is five files. There is no dashboard to learn and no dependency to install beyond Python's standard library; the runner uses urllib and nothing else.
| File | Contents | Why it exists |
|---|---|---|
| prompts.jsonl | 40 objects, one per line | Loadable by any language; no parsing of Markdown |
| run_evals.py | Runner with retry and pacing | Proves the prompts work before you adapt them |
| README.md | How to run, how to edit, model notes | The instructions live next to the files |
| LICENSE.txt | Commercial use, no resale | You know what you bought |
Each JSONL object carries the prompt, the output shape, two or more eval cases, a model recommendation and a note about the failure mode to watch. The note matters as much as the prompt: it tells you what usually goes wrong with that task.
The four categories
Ten prompts each, chosen because they are the tasks teams actually ship first. Every prompt asks for JSON where the task allows it, because JSON is checkable and prose is not.
| Category | Prompts | Example | Default model |
|---|---|---|---|
| Extraction | 10 | Invoice fields, contract renewal dates, meeting action items | plugsky-lite, plugsky-pro for contracts |
| Classification | 10 | Intent router, refund eligibility, content safety | plugsky-lite, plugsky-pro for policy |
| Support | 10 | First reply, escalation summary, churn save | plugsky-lite, plugsky-pro for disputes |
| Coding | 10 | Code review, unit tests, reversible migration | plugsky-coder |
The support prompts use a constraint that is unusual in prompt libraries: several of them forbid specific words. A refund reply may not contain a promise; a churn reply may not invent a discount; a password-reset reply may not ask for a secret. Those prohibitions are eval assertions, not style advice.
How the evals work
An eval is a small assertion set attached to one input. The runner sends the prompt and the input, then checks the output. Five assertion types cover most of what is worth automating.
| Assertion | What it checks | When to use it |
|---|---|---|
json_keys | The parsed output contains every named key | Any structured extraction |
equals | The named field has an exact value | Numbers, enums, booleans |
length | An array has the stated number of items | Action items, line items |
must_contain | The text contains each phrase, case-insensitive | Required facts, policy references |
must_not_contain | The text contains none of the phrases | Promises, discounts, leaked internals |
The runner retries a 429 after the cooldown it reports and paces requests, because a suite of 83 calls will occasionally meet a busy upstream. It exits non-zero if anything fails, so it can sit in CI unchanged.
A worked example
The intent router is a good example of why the eval is the contract. The prompt defines five intents and four rules: a charge on an existing account is billing; buying, seats or contracts before purchase is sales; a failure is technical support; login or team changes are account. The first version of the prompt failed its own third eval because “annual invoicing for 40 seats” was classified as billing. Adding the rule fixed it, and the eval now passes 3/3.
That is the loop the pack is built for: write the rule, run the cases, change the prompt until the cases pass, then add a case for the next ambiguity. Without the case, the failure would have shipped.
What the pack does not do
- It does not prove correctness. The evals check shape and a few facts. A valid JSON answer can still be wrong, and a support reply can still be unhelpful.
- It does not include your policy text. Refund and eligibility prompts take the policy in the request; they do not know your rules.
- It does not tune prompts to your data. The prompts are a strong starting point; the evaluation set that matters is the one built from your inputs.
- It is not a hosted product. There is no prompt dashboard and no evaluation UI. It is a ZIP and a runner.
Those limits are the reason the pack is $49 rather than a platform subscription. It removes the blank page and gives you a test harness; it does not replace the work of understanding your own inputs.
Licence and updates
Commercial use inside your organisation and inside products you build is covered. Reselling or republishing the pack as a prompt library, dataset or competing product is not. The licence file is in the ZIP and says the same thing in one page.
The pack is dated. Model behaviour changes, and a prompt that passes today can fail after an upstream upgrade. Re-run the evals after a model change; that is what they are for. A future revision is a separate purchase unless it is a correction to a file in the version you bought.
Who should skip it
If you already have an evaluation set built from your own traffic, this pack adds little. If your only LLM task is a single summarise button, the prompt is the easy part and you do not need 40 of them. The pack earns its price when a team is shipping several LLM features and each one currently starts from a blank prompt box.
Questions
What exactly do I get?
A ZIP with prompts.jsonl (40 prompts and 83 eval cases), run_evals.py, README.md, RUN-LOG.txt and LICENSE.txt.
Which models does it target?
Plugsky-lite for extraction and classification, plugsky-pro for policy and analysis, plugsky-coder for code tasks. The runner accepts --model to override.
Do the evals prove the output is correct?
No. They prove the output has the required shape and a few required facts. A valid JSON answer can still be wrong; you still review samples.
Can I use the prompts commercially?
Yes, inside your organisation and in products you build. You may not resell or republish the pack itself.
Is it a subscription?
No. It is a $49 one-time purchase; there is no subscription and no renewal.
Sources
- Plugsky pricing and fair-use limits — checked 22 November 2026
- Plugsky model catalogue — checked 22 November 2026
- OpenAI structured outputs guide — checked 22 November 2026
- Stripe Checkout documentation — checked 22 November 2026
- OpenAI chat completions reference — checked 22 November 2026
Related add-ons
Prompt Cost Linter
Static analysis for seven LLM cost anti-patterns: full context resent every turn, missing prompt caching, flagship models on simple tasks, unbounded output, retry storms, batchable loops and needless re-embedding.
Usage Export
Download every request your workspace makes — model, tokens, cost, status and latency — as a CSV with a fixed header. Built for finance reconciliation and warehouse loads.
Stripe checkout. The download appears on this page and in Dashboard → Billing → Add-ons as soon as payment completes.
$49 — buy onceCompare add-ons