Agent Prompt Pack

What's inside

Most prompt packs hand you text and wish you luck. The prompt is the easy part. What's missing is the context: when to reach for it, how to tell whether the answer is any good, and the specific way the obvious version fails. That context is the whole product here — it's why this is 6,200 words for 25 prompts instead of 2,700.

Every prompt ships with four things:

• When — the moment you reach for it
• Expected output — what good looks like, so you can tell it worked
• Failure mode — the specific way the naive version goes wrong
• Tip — the non-obvious thing that makes it land

The failure modes are real, not hypothetical. They come from running a 42-agent system with 368 scheduled jobs on one Mac at $0/month in API costs. A sample of what's in there:

• A health check that read a config field and reported healthy while the browser held zero session cookies and could not load a single authenticated page. It was reading a record of intent and reporting it as fact.
• except Exception that did not catch the exception — a safety gate called sys.exit(), and SystemExit inherits from BaseException. The guard meant to make failure safe was killing the caller instead.
• A maintenance script whose every health branch was silently "command not found", so it declared "browser down — restarting" on every run while the browser was fine.
• A write that passed the lock, the file hash and the journal and was still wrong, because all three watch bytes and staleness is semantic.

Also inside: an architecture diagram of the loop every agent lives in, a table mapping each layer to its prompts and its most common failure, a worked example showing a real input with its real output (and what a failing output looks like, so you can recognise one), and 15 links to genuinely free tools and docs — every URL fetched and confirmed working before publication.

What this is not: a course, a framework, or a codebase. There is nothing to install. If you want copy-paste code that runs an agent today, this is the wrong purchase — this is design-level material. And if you have already shipped several agents to production, you will likely know a third of these failure modes already.

Costs nothing to run. Every prompt works against a local model (Ollama) or a free cloud tier. No paid API key at any point.

Try before you pay: the 7-prompt starter set from the same system is free — search "7 Prompts From a Live Autonomous Agent Stack". If that one earns its keep, this is the full set.

25 prompts · 100 annotations · ~6,200 words · markdown · yours forever, updates included.