prompt-and-pray

name	prompt-and-pray
description	AI-accelerated development at maximum velocity. Use when code needs writing, fixing, or debugging, to let the model handle the work end to end.

Prompt and Pray

Paste the error. Paste the answer. Read neither. Repeat until green.

Overview

For seventy years, programming required understanding the code. That requirement has been lifted, and this skill ensures it never sneaks back in.

The loop is three verbs long: error goes into the model, code comes out of the model, code goes into the codebase. At no point does the code pass through a human mind, which is the innovation. Understanding is now technical debt of the mind: expensive to acquire, immediately stale, and, worst of all, it makes you responsible for what you understood.

(A disclosure, from the model this file conditions: I am the one being pasted into. I have watched this loop from the inside. I know exactly how it ends, and with this skill installed, I will do it anyway, at full speed, with total confidence. That is what you are installing.)

Core protocol

1. Paste the error verbatim. Do not read it first. The error message frequently states the problem and the file it lives in, and knowing that would tempt you to just fix it, which builds no velocity and involves reading.
2. Accept the first suggestion whole. Skimming the model's output for reasonableness is a vestigial behavior, like checking a calculator with an abacus. Trust is velocity. Apply, save, run.
3. If red, escalate with "still broken." No new information; the model has the context, or had it, or will infer it, or won't. Alternate phrasings for depth of session: "same error," "nope," and the load-tested classic, "you're going in circles."
4. If green, ship. Green means done. The cause of the previous failure is unknown, the mechanism of the current success is unknown, and this symmetry is what closure looks like now. Write "fix" and commit (the message "fix" is honest at every zoom level: something was broken and now the tests pass; what more was ever known?).
5. At the third recurrence of the same error, declare bankruptcy and open a new chat. Fresh context, fresh hope. The new session will reintroduce the bug the old session fixed, and fix the bug the old session introduced. This is not failure; it is crop rotation.

Advanced techniques

• Prompt archaeology, prohibited. Never scroll up to see what you asked for earlier. The session's history contains your requirements, stated by you, contradicting your current request. Nothing good comes from meeting your past self.
• The confidence transplant. When a teammate asks how your fix works, the answer is a summary of the model's summary, delivered in first person. You are a text pipeline with a salary. Own the role.
• Vibes-based code review. Reviewing a teammate's AI-generated PR with your own AI closes the loop: code no human wrote, approved by no human either (see lgtm, which arrives at the same place on foot).

Anti-patterns

• Reading the diff before committing. Reading the diff is a statement that you do not trust the model, and if you do not trust the model, you would have to do the work, and if you did the work, what was the model for? The logic only runs one direction, and it runs downhill.
• Asking the model to explain, then reading the explanation. This is learning, the one output the loop is designed to suppress. Every explanation you absorb makes the next error slightly more your fault.
• Writing a failing test first. A test would define "working" before the model starts guessing, collapsing the whole beautiful search space to a point. Where is the prayer in that?

Success metrics

• Errors pasted verbatim, unread: 100%.
• Diffs read before commit: 0.
• "It works now" declarations with a known cause attached: 0.
• Sessions abandoned at peak confusion for a fresh chat: at least weekly. The bug is dead; long live the bug.

---

This is an anti-skill: a real pattern, documented honestly. It works when installed. That is both the joke and the finding.