bro-skill is thirty lines of prompt that refuse to add a single fact

One markdown file, one job, no new information

The entire skill is `SKILL.md`, and the project describes it as a 30-line prompt file. You type `/bro` after a reply that was too dense, too jargon-heavy or too formal, and the assistant re-explains its own previous message as a friend would over a beer.

The constraint that defines the design is stated first: no new information and no new answers, just the same content said in a way that cannot be misunderstood. That is a narrow target and the reason the file is small. A rephrase feature has an obvious failure mode, where the model helpfully adds the missing detail you did not ask for, and this one is written specifically to avoid it.

What you get instead of a new answer is a re-rendering of what the model already said. Compared with typing please explain that again more simply into every conversation, the difference is that the constraints are written down once: no new facts, verbatim commands, same language, flattened structure. Compared with handing the text to a separate summarisation tool, the difference is that it operates on the conversation's own last message rather than on input you paste.

Facts survive verbatim, and tools are off limits

Two of the six rules exist to keep the rephrase from lying. The first says every path, command, filename, number and URL stays exactly the same. The second says the skill never answers a new question, never adds information and never calls tools.

That combination is stricter than it first sounds. A shell transcript is exactly the kind of text a simplifier wants to tidy, and the repository keeps a dedicated example file for that case so a reader can watch commands come through untouched. It is named for watching commands survive verbatim, which tells you what the author was worried about.

The no-tools rule matters as much. Because the skill cannot fetch anything, it cannot rescue a reply that was wrong or out of date, and it cannot quietly look something up to patch a gap. That is the price of the guarantee, and it is the right trade for a command you reach for reflexively, since a rephrase button that silently improves the content would teach you to trust it more than a fixed prompt deserves.

Same language in, same language out

The rule stated as PT-BR in, PT-BR out is the one that distinguishes this from a generic simplify command. English stays English, and Portuguese stays Portuguese, which matters because the input and the reader's fluency are not the same thing.

The example makes the point concretely. A passage of consultant Portuguese about conversion funnel optimisation, omnichannel nurturing and lifetime value comes back as Portuguese, keeping the arithmetic intact in plain words: fewer customers leaving plus each one spending more equals more money, which is what the acronym soup was claiming.

The companion rule is simpler, not shorter, meaning clarity takes priority over word count and the rewrite is allowed whatever space it needs. That is a deliberate rejection of the instinct to compress, and it is the difference between a rephrase you can follow and one that has merely been trimmed. It also means the output can occasionally be longer than the input, which is the correct outcome for a dense paragraph that hid its point in subordinate clauses.

Flatten structure and add a light bro flavor

Headers and tables become plain sentences, which is the rule with the most visible effect on the output. A structured answer with a summary table reads as authoritative and unforgiving; the same content as running prose reads as an explanation, which is what someone who just said bro what needs.

The architecture example shows the transformation. A sentence about an idempotent reconciliation layer over an event-sourced aggregate, mitigating dual-write anomalies during a migration window while preserving the legacy projection contract, comes back as: a safety layer that keeps old and new data in sync while the systems are being switched, with nothing breaking for anything still on the old one.

The last rule sets the register, described as light bro flavor: casual and direct, not a meme. That boundary is doing real work. A skill that overdoes the persona turns every explanation into a performance, and a skill that does not lighten the register at all has not solved the complaint. The Kubernetes example shows the intended middle, where the assistant keeps the actionable detail, naming `initialDelaySeconds` and suggesting an init container, while dropping the passive construction that made the original hard to read.

Five tools, five directories, and one folder Cursor does not have

The skill targets the Agent Skills format and ships thin slash-command wrappers where a skill alone does not give you a typed command. Each tool wants the file in a different place:

| Tool | Skill location | Typed command via | |---|---|---| | Hermes Agent | ~/.hermes/skills/bro/SKILL.md | skill, automatic | | Claude Code | ~/.claude/skills/bro/SKILL.md | ~/.claude/commands/bro.md | | Cursor | ~/.cursor/skills/bro/SKILL.md | skill, typed directly | | OpenAI Codex | ~/.codex/skills/bro/SKILL.md | ~/.codex/prompts/bro.md | | opencode | ~/.config/opencode/skills/bro/SKILL.md | ~/.config/opencode/command(s)/bro.md |

The Cursor row has a footnote worth reading twice. Cursor also reads the Codex and Claude skill directories as a compatibility fallback, and the README says not to rely on it, because a machine that only has Cursor often has neither of those directories. So the installer writes `~/.cursor/skills/bro/` whenever it detects Cursor, creating `~/.cursor/skills/` in the process, since that folder is not there by default.

Install is a one-liner, a clone, or --all

The short path detects which agent CLIs you actually have and installs into each:

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

Cloning is the better route if you want updates later:

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

The difference between the two flags is scope. Detection only touches tools present on the machine, while `--all` writes into every supported tool whether or not you use it, which on a fresh account means creating five sets of directories. Updating is a pull and a re-run, or simply re-running the one-liner, which is described as idempotent:

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

Either way the session has to be restarted before `/bro` is typed, since agents read their skill directory at startup.

What /bro cannot do for you

The limitation is the same one as the feature, seen from the other side. Because the skill never calls tools and never adds information, a reply that was wrong stays wrong in plainer words, and a reply that omitted something stays incomplete. That is the correct behaviour for a command whose promise is faithfulness, but it means `/bro` is a reading aid rather than a second opinion.

The other constraint is where the files land. Installing writes into your home directory across up to five tool configurations, creates a Cursor skills folder that does not normally exist, and the one-liner executes a script fetched over the network. All of that is proportionate to a 30-line markdown file, and all of it is worth doing with your eyes open rather than through muscle memory.

The repository is small and stable: `SKILL.md`, `install.sh`, `examples/`, the README and an MIT licence, with five example files covering consultant speak, a git panic, a Kubernetes exchange, a Portuguese case and a software architecture case. It is not archived, the last push was on 2026-08-27, and there are no tagged releases, so there is no version to pin and updates arrive by pulling the branch.

Editorial conclusion

bro-skill suits anyone who reads dense model output daily and wants a fixed rephrase button instead of retyping a request, and it is a good template if you are writing your own Agent Skill because the rules are short enough to hold in your head. It does not suit anyone who wants a second opinion, since a wrong previous reply is faithfully restated rather than checked, and it cannot look anything up because it never calls tools. Verify first which of the five tool directories the installer will create on your machine, whether you want `--all` writing into tools you do not use, and whether your agent reloads skills without a session restart.

Frequently asked questions

Official sources

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