| name | buddy |
| description | REQUIRED when the user explicitly says "buddy", "task-buddy", or "verify-buddy" (or plurals "buddies", "task-buddies", "verify-buddies", and close variants like "use buddy", "run a task-buddy", "do a verify-buddy", "run 4 verify-buddies"). Do not trigger on generic phrases like "second opinion", "external review", "another model", or "GPT" -- only on the literal terms buddy / task-buddy / verify-buddy (and their plurals). |
| allowed-tools | Bash(buddy *), Bash(vdiff *), Agent, Read |
Buddy
buddy is a read-only wrapper around pi -p running an external
model. Use it for second opinions, never as the primary author.
Mode selection
The user's word picks the mode. Do not switch modes on your own.
| User says | Mode | What it means |
|---|
| "buddy" | Direct | Claude runs buddy itself. |
| "task-buddy" | task-buddy | Subagent passes the request to buddy and returns a double-checked report. |
| "verify-buddy" | verify-buddy | Subagent does the work, has buddy double-check, and returns the report. |
Rules (every mode)
- No secrets, credentials,
.env, customer data, or production logs.
- Narrow paths only.
@path is a pi context ref, not a glob -- no
directory expansion. Avoid vendored dirs and large diffs; split into
focused calls.
- Advisory only. Verify file paths, symbols, and line ranges against
the source before promoting any finding -- buddy hallucinates refs.
- No session. Every call is self-contained; put all context in one
prompt.
- Buddy tools are
read,grep,find,ls only. It cannot run tests, VCS,
or builds.
- Subagents: no edits, no commits, no installs.
Invocation
buddy @path/to/file.ts "focused review prompt"
buddy @a.ts @b.ts "prompt"
vdiff | buddy "review this diff"
buddy @file.ts <<'PROMPT'
multiline prompt
PROMPT
Stdin and args both feed pi -- pick one cleanly, do not rely on merge
order.
Goals per mode
Buddy hallucinates refs, so double-checking is part of the goal in
every mode. For task-buddy and verify-buddy, the spawned subagent
does not see this skill or your conversation -- hand over goals and
the user's request, not steps.
buddy (direct)
- Pass the user's request with relevant context to buddy.
- Double-check buddy's report against the source.
- Respond with the double-checked report.
task-buddy
Goals for the subagent:
- Pass the user's request with relevant context to buddy.
- Double-check buddy's report against the source.
- Return the double-checked report.
Example, given user said "task-buddy review plan//guide.md
for implementation readiness":
The user asked: "review plan/<feature>/guide.md for
implementation readiness -- unanswered questions, design bugs,
anything needing further planning."
Run buddy with the relevant context, double-check buddy's
report against the guide, and return the double-checked report.
Do not edit, commit, or install.
verify-buddy
Goals for the subagent:
- Do the user's request.
- Ask buddy to double-check your findings.
- Return the double-checked report.
Example, given user said "verify-buddy review plan//guide.md
for implementation readiness":
The user asked: "review plan/<feature>/guide.md for
implementation readiness -- unanswered questions, design bugs,
anything needing further planning."
Do the work, ask buddy to double-check your findings, and
return the double-checked report. Do not edit, commit, or
install.
Reporting
Report what is useful for the task -- format is yours to choose. Call
out where you disagree with buddy.