luchasarie/bro-skill: a one-file /bro command that re-explains the last answer

The problem: you asked a question and got a wall of jargon back

Coding agents write for the transcript, not for the reader. A refactor explanation comes back as "an idempotent reconciliation layer atop the event-sourced aggregate", which is technically accurate and useless when you are trying to decide whether to merge. The usual fix is to type "explain that more simply", which starts a new turn, invites the model to add detail, and often produces a third paragraph of the same vocabulary.

bro-skill takes a narrower position. It is a skill whose only job is to restate the assistant's own previous message in plain language, with no new information and no new answers. The README frames it as "like a smart friend over a beer". The audience is anyone who already uses an agent for work and occasionally loses the thread of what it just said, including non-native English speakers, since the skill preserves the language of the original reply.

What SKILL.md actually constrains the model to do

The repository is five entries: LICENSE, README.md, SKILL.md, examples/ and install.sh. SKILL.md is the entire skill, and the README describes it as a 30-line prompt file. The behaviour lives in a table of rules rather than in code, which matters because it means there is no parser, no runtime and nothing to debug beyond the prompt text.

The rules are: re-explain, don't re-answer; simpler, not shorter; facts survive verbatim; light bro flavor; same language; flatten structure. The first rule is the load-bearing one, because it is what stops the skill from drifting into a general-purpose assistant. The third rule is the one with real teeth for engineering work: every path, command, filename, number and URL must stay exactly the same. The examples directory includes git-panic.md, which the README points at specifically so you can watch the commands survive verbatim. The last rule converts headers and tables into plain sentences, which is why a /bro reply reads as prose rather than a reformatted version of the original.

The same-language rule is worth calling out because it is easy to get wrong in a prompt. The README gives a PT-BR example where the dense input and the plain output are both Portuguese. English stays English. There is no translation feature and the examples/pt-br.md file exists to pin that behaviour down.

Installing bro-skill and getting a first /bro out of it

The README offers a one-liner that detects which agent CLIs you have and installs into each of them. Running it is the fastest path, and the script is described as idempotent, so re-running it later is safe.

bash <(curl -fsSL https://raw.githubusercontent.com/luchasarie/bro-skill/main/install.sh)

If you would rather keep a local copy so you can pull updates, the README gives the clone route instead. The plain ./install.sh installs into every detected tool, and --all forces installation into every supported tool rather than only the detected ones.

git clone https://github.com/luchasarie/bro-skill.git
cd bro-skill
./install.sh          # installs into every detected tool
./install.sh --all    # or force-install into every supported tool

After either route, the README says to restart your agent session, then type /bro after any reply. Where the file lands depends on the tool: ~/.hermes/skills/bro/SKILL.md for Hermes Agent, ~/.claude/skills/bro/SKILL.md for Claude Code, ~/.cursor/skills/bro/SKILL.md for Cursor, ~/.codex/skills/bro/SKILL.md for Codex, and ~/.config/opencode/skills/bro/SKILL.md for opencode. The README notes that ~/.cursor/skills/ is not present by default and that the installer creates it when Cursor is detected.

A skill alone does not always give you a typed slash command, which is why the repository ships thin wrappers. Claude Code gets ~/.claude/commands/bro.md, Codex gets ~/.codex/prompts/bro.md, and opencode gets ~/.config/opencode/command(s)/bro.md. Hermes and Cursor are listed as exposing /bro through the skill itself. If typing /bro does nothing after a restart, the wrapper path for your tool is the first thing to check.

The limits: no summarising, no new answers, no tool calls

The design deliberately removes capability. The skill never answers a new question, never adds information and never calls tools. That means it cannot look up the Kubernetes documentation to explain CrashLoopBackOff better than the original reply did; it can only rephrase what is already on screen. If the first answer was wrong, /bro will produce a clearer version of the same wrong answer.

"Simpler, not shorter" is also a constraint people misread. The README states clarity over word count, so a /bro reply can be longer than the dense paragraph it replaces. If you were hoping for a one-line summary, this is the wrong tool, and the skill is explicitly not a summariser.

The verbatim rule cuts the other way too. Paths, commands, filenames, numbers and URLs have to survive unchanged, so /bro will not clean up a malformed command or fix a wrong flag. It also will not translate: a Portuguese reply stays Portuguese, an English reply stays English. And the flavour is bounded by the prompt rather than by the model's taste, so it is a fixed register rather than something you tune per project. Anyone who wants adjustable tone should edit SKILL.md directly, since that file is the whole implementation.

How this differs from asking the agent to explain again

The obvious alternative is the one everybody already uses: type "explain that more simply" in the same session. The difference is scope. A fresh instruction is a new turn, and a new turn is an invitation to add detail, restructure the answer, or drift into the underlying question again. bro-skill is a persisted rule set that constrains the reply to a re-explanation of the previous message, with the no-new-information rule stated up front.

The second alternative is a general summarisation prompt or a separate summariser tool. Those compress, which is the opposite of what the skill's "simpler, not shorter" rule asks for, and they typically do not carry a verbatim-facts guarantee. For a reply containing a kubectl command or a file path, compression is where information gets lost.

The honest comparison is that bro-skill is not competing on capability. It is a 30-line prompt file with a narrow contract, and its value is that the contract is written down and portable across five tools rather than retyped per session.

Updating, licence and the cost of keeping it around

Updating is a git pull plus a re-run of the installer, or simply re-running the one-liner, which the README describes as idempotent.

cd bro-skill && git pull && ./install.sh

There are no releases listed for the repository, so updates arrive as commits on main rather than as tagged versions. That is a maintenance consideration rather than a defect: you cannot pin a version, and a pull takes whatever is on the branch. The last push was on 2026-08-27, so the project has moved recently, but with no release artifacts there is nothing to compare against.

The repository is MIT licensed, which permits reuse and modification subject to the licence terms; keep the LICENSE file with any copy you redistribute. Because the skill is a prompt rather than a dependency, the upgrade cost is bounded by how much you have customised SKILL.md. If you have edited it, a pull can conflict, and the README's contributing note suggests the maintainer treats the phrasing as the sensitive part by asking contributors to keep the facts-verbatim rule intact.

Where the documentation is thin

The README does not document rollback or uninstall. There is no command to remove the skill from the five tool directories, and no list of what the installer touches beyond the skill and wrapper paths in the compatibility table. If you run ./install.sh --all on a machine where you only wanted one tool configured, you will be cleaning up by hand.

Detection behaviour is also described only in outline. The README says the script detects which tools you have and that Cursor causes ~/.cursor/skills/ to be created, but it does not say how detection works or what happens when a tool's directory exists but the CLI does not. The compatibility table notes that Cursor reads ~/.codex/skills/ and ~/.claude/skills/ as a fallback, and the README advises against relying on it because Cursor-only machines often have neither directory. That is a useful warning, but it also means the effective install location for a given machine is something you verify rather than assume.

Editorial conclusion

Adopt bro-skill if you run Hermes, Claude Code, Cursor, Codex or opencode and regularly skim past dense agent replies; the install script writes into the skill directory of each tool it detects, and the whole behaviour is one readable markdown file you can edit. Skip it if you want summarisation, translation, or a second opinion on the original question, because the skill is explicitly forbidden from adding information or calling tools. Before relying on it, open SKILL.md and confirm the same-language rule and the verbatim-facts rule, then check that the wrapper file exists at the path your tool expects, since a skill alone does not always give you a typed /bro.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. luchasarie/bro-skill on GitHub
4. README