一键导入
talk-to-claude
Talk to Claude asynchronously about a specific task by asking it to write its response into a file, then wait for that file to appear.
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。
菜单
Talk to Claude asynchronously about a specific task by asking it to write its response into a file, then wait for that file to appear.
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。
基于 SOC 职业分类
Create or refactor Sifr demos under demos/ when the task is to add a runnable demo, rename a demo, simplify an existing demo, preserve demo history, or keep demos feature-first instead of process-, planning-, or problem-solution driven. Use for demo naming, history headers, compact assert-driven examples, and moving non-demo artifacts out of demos/.
End-to-end execution loop for ad-hoc phases in Sifr: per-wave implementation, validation, PR/merge, external review pass-1/pass-2, then wave/milestone/phase closure cycles with the same review loop, notifications, and final closure bookkeeping.
Documents the sprint-based project workflow for the sifr GitHub project. Covers the full ticket lifecycle from planning through review, ticket types (Epics and Tasks), PRDS documents, board columns, and which commands to use at each phase. Use when the user asks about the workflow, sprint process, ticket management, or needs guidance on which command to run next.
| name | talk-to-claude |
| description | Talk to Claude asynchronously about a specific task by asking it to write its response into a file, then wait for that file to appear. |
Choose a target file under ./plans/reviews/active, then run Claude CLI in the background and redirect its output there.
Use an absolute path derived from ./plans/reviews/active so the output location is exact.
If you need the wait helper, set TALK_TO_CLAUDE_PROJECT to the local
talk-to-claude checkout. Do not hard-code a personal absolute path.
Authoritative launch command:
PWD_NOW="$(pwd)"
mkdir -p "${PWD_NOW}/plans/reviews/active"
TARGET_FILE="${PWD_NOW}/plans/reviews/active/<conversation-file-name>.md"
LOG_FILE="${TARGET_FILE%.md}.claude.log"
nohup claude --dangerously-skip-permissions --setting-sources project --model claude-opus-4-7 --effort xhigh -p "$(cat <<PROMPT
${TOPIC_NAME}
PROMPT
)" >"${TARGET_FILE}" 2>"${LOG_FILE}" &
CLAUDE_PID=$!
wait "${CLAUDE_PID}"
CLAUDE_EXIT_CODE=$?
if [ "${CLAUDE_EXIT_CODE}" -ne 0 ] || [ ! -s "${TARGET_FILE}" ]; then
echo "Claude reviewer failed or produced an empty output. See ${LOG_FILE}" >&2
exit 1
fi
<conversation-file-name>.md with a concrete file name for the active topic.wait "${CLAUDE_PID}" in the same shell invocation; detached nohup ... & can leave empty artifacts in Codex exec sessions.Use the wait command below to block until Claude writes the target file:
PWD_NOW="$(pwd)"
: "${TALK_TO_CLAUDE_PROJECT:?Set TALK_TO_CLAUDE_PROJECT to the talk-to-claude checkout path}"
uv run --project "${TALK_TO_CLAUDE_PROJECT}" \
python "${TALK_TO_CLAUDE_PROJECT}/wait_for_review.py" "${PWD_NOW}/plans/reviews/active/<conversation-file-name>.md" \
--timeout-seconds 2400 \
--poll-seconds 10
If Claude does not produce the file, inspect the per-conversation log:
PWD_NOW="$(pwd)"
sed -n '1,200p' "${PWD_NOW}/plans/reviews/active/<conversation-file-name>.claude.log"
Read the review file, separate actionable findings from non-blocking suggestions, and act on the findings according to engineering judgment. If a recommendation is not worth taking, explain why briefly.