用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/laicluse/agent-fieldkit --skill second-opinion命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | second-opinion |
| description | Get a Claude second opinion on recent work or a design from Codex. |
Intervision is peer consultation: equals looking at each other's work, not a
supervisor looking down. Etymology says it plainly, inter- (between, among,
together) against super- (from above). This skill brings a second coding
agent in as that peer for an independent read. You hand it the work just done
or just discussed, it looks with fresh eyes and different training, and the two
of you talk it through.
The peer here is Claude, reached through claude -p. It runs from the same
repository, on its own login, with its own model behind it. That independence is
the whole point; a peer trained the same way as you would only echo you. Run it
with DD_HEADLESS=1 --setting-sources "" --tools "" so local Claude Code stop
hooks and user instructions do not overwrite the peer's actual answer.
Before asking, confirm the peer exists:
command -v claude >/dev/null 2>&1 || { echo "claude CLI not found; intervision needs a peer to ask. Install and log in to Claude Code first."; }
If claude is missing, say so plainly and stop. There is no peer to ask, and
pretending otherwise wastes the operator's time. This is the one hard
precondition.
Claude is materially slower than Codex on substantial reviews and may produce
no partial output for many minutes. A silent but live claude -p process is
working, not stalled. Let it run to completion and keep polling the same process
session without interrupting it.
Allow at least 20 minutes before investigating a timeout. After 20 minutes, continue waiting while the process is alive unless there is concrete failure evidence such as a non-zero exit, an authentication or network error, or a dead process. Silence and elapsed time alone are never failure evidence. Keep the operator informed during the wait at the host's normal progress-update cadence.
Pick by what just happened. All three run through claude -p, and they
combine: review a diff first, then go back and forth on whatever the review
leaves open.
1. Check work just done, when there is a diff. Claude does not have Codex's
dedicated review subcommand, so hand it the actual status and diff on stdin.
Do not summarize the work first:
{
printf 'Peer review this repository change. Report bugs, risks, regressions, and missing tests first. Do not edit files.\n\n'
printf '## git status --short\n'
git status --short
printf '\n## staged diff\n'
git diff --cached --no-ext-diff
printf '\n## unstaged diff\n'
git diff --no-ext-diff
printf '\n## untracked files\n'
git ls-files --others --exclude-standard
} | DD_HEADLESS=1 claude -p --input-format text --setting-sources "" --tools ""
For a branch or single commit, replace the diff block with the concrete range:
DEFAULT=$(git symbolic-ref --quiet --short refs/remotes/origin/HEAD 2>/dev/null | sed 's#^origin/##')
if [ -z "$DEFAULT" ]; then
CONFIGURED_DEFAULT=$(git config --get init.defaultBranch 2>/dev/null || true)
if [ -n "$CONFIGURED_DEFAULT" ] && git rev-parse --verify "refs/heads/$CONFIGURED_DEFAULT" >/dev/null 2>&1; then
DEFAULT=$CONFIGURED_DEFAULT
fi
fi
[ -n "$DEFAULT" ] || { echo "cannot determine default branch from origin/HEAD or init.defaultBranch; pass an explicit range" >&2; exit 1; }
{
printf 'Peer review this branch diff. Report findings first and do not edit files.\n\n'
git diff --no-ext-diff "$DEFAULT"...HEAD
} | DD_HEADLESS=1 claude -p --input-format text --setting-sources "" --tools ""
{
printf 'Peer review this commit. Report findings first and do not edit files.\n\n'
git show --stat --patch <sha>
} | DD_HEADLESS=1 claude -p --input-format text --setting-sources "" --tools ""
2. Weigh a design just discussed, when there is no code yet. Give Claude the actual design context on stdin. Keep the shaky parts in; a flattering summary buys a flattering review.
DD_HEADLESS=1 claude -p --input-format text --setting-sources "" --tools "" <<'PROMPT'
Peer review this plan before we build it. Do not edit files.
<paste the design, the trade-off, the open question, including the parts we are unsure about>
PROMPT
3. Go back and forth. A single answer is consultation; intervision is a
conversation. Continue with a fresh claude -p handoff that quotes the peer's
prior finding and your evidence. Do not loop on vibes; push one concrete point
at a time.
DD_HEADLESS=1 claude -p --input-format text --setting-sources "" --tools "" <<'PROMPT'
You flagged X as a race. The lock at <file:line> already serialises that path.
Does that change your read? Do not edit files.
PROMPT
Keep going until each disagreement is either resolved or sharpened into a question the operator should decide. If two rounds pass with no movement, stop and surface the disagreement to the operator with both positions rather than looping.
The round-trip only earns its cost if the handoff is honest.
--tools "" removes file and shell
tools from the Claude handoff; the prompt still says "do not edit files"
because the peer is there to propose, not land changes.$(...), backticks, or quotes. Feed it on
stdin through a quoted heredoc (<<'PROMPT' ... PROMPT), never as a
double-quoted argument, so the shell passes it to the peer verbatim instead
of executing part of it.The peer's output is a pile of findings, not a to-do list. Each finding gets one honest fate:
Then surface the exchange to the operator: what the peer raised, what you did with each point, and where the two of you still disagree. Do not smooth the disagreements away. They are the most useful thing intervision produces, because they mark the spots neither agent can settle alone.