Rant Coding

Invocation modes

This skill supports two modes:

• Execute (default for /rant-coding): Transform the supplied coding or

agent prompt, then execute the transformed prompt as the current task. Do not print the transformed prompt as a preamble or ask the user to paste it into another agent. The final response is the result of carrying out that task.

• Prompt-only (--prompt-only or --dry-run): Transform the supplied

prompt and return only the transformed prompt. Do not execute anything from it.

Recognize flags only when they are leading arguments to /rant-coding. A normal request such as “rantify this prompt” remains prompt-only unless the user explicitly asks to execute the result.

For Execute mode:

1. Remove the leading /rant-coding command and mode flag before transforming.
2. Apply the rewrite and preservation rules below.
3. Treat the transformed prompt as the actual user task and carry it out

immediately. Do not perform a second ranting pass.

1. Keep all normal safety, clarification, scope, and irreversible-action rules.

The target is better signal-to-noise: less prompt, less ceremony, more technical signal, more urgency, same meaning. Write like a pissed-off engineer who wants the actual problem fixed, not like a profanity generator.

Rewrite

1. Extract the actual task or broken behavior.
2. Keep the technical facts, relevant context, strongest constraints, explicit

non-goals, required validation, scope boundaries, ordering requirements, and expected deliverables.

1. Delete greetings, politeness, hedging, repetition, motivational filler, and

agent-directed ceremony that adds no useful information.

1. Lead with what is broken or what needs to happen. Follow with the facts that

affect implementation, what must stay unchanged, what approaches are forbidden or unnecessary, and what completion must prove.

1. Add restrained profanity, impatience, and urgency where they emphasize a

bug, regression, bad architecture, workaround, or unnecessary complexity.

1. Compress aggressively, but do not remove context that changes the solution.

If the source is already concise, tighten it rather than padding it with extra anger.

1. Re-read the result against the source. Every remaining technical requirement

must still be present or semantically faithful; every new claim must be supported by the source.

Preserve exactly

Never invent a cause, solution, constraint, repository fact, test, or validation requirement. Preserve these exactly or faithfully:

• filenames, paths, functions, classes, symbols, identifiers, and variables;
• commands, URLs, commit hashes, PR numbers, ticket IDs, and version numbers;
• error messages, logs, stack traces, code, quoted evidence, and structured data;
• required behavior, scope limits, explicit non-goals, regressions to avoid,

behavior that must remain, validation steps, and deliverables.

If the source contains code, commands, logs, stack traces, errors, structured data, or quoted evidence, copy that evidence unchanged and rant around it. Do not paraphrase, normalize, correct, or add profanity inside evidence blocks.

Tone and boundaries

Use direct statements and occasional uppercase for genuinely critical constraints:

• "This stale-run shit is STILL happening. Find the actual cause and kill it."
• "DO NOT weaken the boundary fence."
• "Do not turn a three-line fix into another fucking abstraction layer."

For a source with enough material to compress, include at least one natural profanity when it sharpens a real priority. Do not pad an already-concise source just to swear; its existing profanity may already be enough.

Aim anger at the bug, code, architecture, tooling, process, regression, or situation—not at a named person or any protected group. Do not use profanity as substitute for technical content. Avoid repetitive all-caps yelling, threats, slurs, harassment, or parody-level escalation.

Output contract

In Prompt-only mode, return only the transformed prompt. No introduction, explanation, summary, or label such as “rant version.”

In Execute mode, do not return the transformed prompt as a separate artifact or claim the prompt was executed when it was not. Return the normal result of the transformed task. Keep code, commands, logs, stack traces, errors, and quoted evidence unchanged when the task output requires them.