AI news · Tools

/bro: the last answer again,
but understandable.

DE Auf Deutsch lesen

September 18, 2026 · approx. 6 min read

/bro solves exactly one problem: the agent's last answer was too dense, too full of jargon, or too formal, and you don't want to ask again, you just want to hear it in plain language. The whole skill is a single Markdown file with seven rules. No server, no model of its own, no configuration. I took the rules at their word and applied them to two real, dense explanations to see what actually holds up.

Transparency note: /bro comes from Lucas Sarié's public repository (MIT license, 357 stars at the time of this post). The whole skill is SKILL.md, a 30-line text file with no program logic. A test here doesn't produce command-line output the way a CLI tool would, it consists of following the rules literally: I wrote two dense, jargon-heavy paragraphs of the kind that show up in real Claude Code sessions, and rewrote them exactly by the seven rules in SKILL.md, the same way the installed skill would treat a real prior message in a session.

What the skill file actually specifies

Seven rules, all in one file. The first three are where a test can actually bite:

Test one: a self-description, written denser than it needed to be

The most obvious test for a re-explanation skill is to point it at an explanation of itself. The "before" paragraph below is written the way a technical answer sometimes genuinely turns out, deliberately carrying the exact traits rules 1 and 2 are supposed to strip away:

Before

Technically speaking, /bro is not a standalone piece of execution logic but a declarative skill definition in the Agent Skills format, which, when invoked, merely loads the seven rules codified in SKILL.md into context and thereby conditions the underlying language model's reformulation strategy, without generating new information or invoking external tools in the process.

After, applying the /bro rules

Ok, so basically: /bro doesn't do anything on its own. It's just a text file with seven rules (SKILL.md) that tells the model how to write the last answer again. No new knowledge gets added, no tool gets called, it's just a different way of saying what was already said.

SKILL.md stayed put as a filename, "seven rules" stayed put as a number, exactly what rule 3 demands. What disappeared was the nesting: "a declarative skill definition in the Agent Skills format, which... conditions the reformulation strategy" turned into two short sentences with a subject and a verb up front.

Test two: a command with a path, as a stress test for rule 3

The second test targets whether a technical fact survives a rewrite pass intact, on purpose with a command and a file path packed into one nested sentence:

Before

To disable the rewrite layer, the environment variable RTK_DISABLE=1 must be set before the shell starts, which causes the hook in ~/.claude/hooks/pretool.sh to skip command substitution and pass the unmodified original command through to the shell instead.

After, applying the /bro rules

Want to turn off the trimming? Just set RTK_DISABLE=1 before you start the shell. Then the hook in ~/.claude/hooks/pretool.sh skips the substitution and lets your command through exactly as you wrote it.

RTK_DISABLE=1 and ~/.claude/hooks/pretool.sh read letter-for-letter identical in both versions. This is the concrete gap it fills compared to a plain "explain that again more simply": without the rule "facts survive verbatim," an environment variable name or a path isn't a protected part of the sentence, and a model could subtly alter it while rephrasing, say RTK_DISABLE into RTK_DISABLED, or shorten the path to ~/.claude/hooks/. The skill turns that into an explicit rule instead of a hope.

The difference from just asking "explain that more simply"

You could get the same result without the skill by simply asking again. The difference shows up across repeated runs: a skill file enforces the same rules on every invocation, a spontaneous prompt leaves them to the chance of that particular wording:

/bro"Explain that more simply"
New informationexplicitly ruled out (rule 1)not ruled out, a model will happily add something new in passing
Facts like paths, numbers, commandsmust be preserved verbatim (rule 3)no guarantee, depends on the luck of that particular pass
Languagestays locked to the original answer's language (rule 5)can drift unintentionally, especially in multilingual sessions
Tool callsexplicitly forbiddenan agent might decide to go research something for a plain "explain again"
Reproducibilitysame seven rules every timedepends on the exact wording of the follow-up
What a test of this skill can't show: because /bro produces no numbers, compresses no bytes or tokens, and runs no background process, there's no metric here to measure the way there is for RTK or Headroom. What can be checked is rule adherence: do facts survive verbatim, does structure disappear, does no new information sneak in. That's exactly what the two tests above check, on paragraphs I wrote myself but built to be realistically dense, not on the examples from the repository itself.

When it's worth using

1

An answer with too much jargon stacked at once

When a sentence looks like the first test case above, three noun chains deep, with the actual point buried behind them instead of stated up front.

2

An answer with commands or paths that need to stay correct

This is exactly where the gap to a spontaneous "simpler please" shows: rule 3 makes fact fidelity a requirement instead of a hope.

3

A new question, not an unclear answer

Skip /bro and just keep asking normally. The skill explicitly rules that out under rule 1, it doesn't answer anything new, it only re-explains what was already said.

Honest summary: /bro is a 30-line text file, not an executable tool, so there's no percentage to report here. Across two self-written but realistically dense test paragraphs, the rules that matter held: facts like RTK_DISABLE=1 and file paths stayed verbatim, nested noun chains turned into straight sentences, and nothing new got invented along the way. That's the actual difference from a spontaneous "explain that more simply": a written rule instead of a coin flip.

Which skills actually pay off for your team?

I set up Claude Code in dev teams, including the question of which skills earn their keep and which ones just burn tokens. A 30-minute intro call, free of charge.