菜单
Track task steps and write reports to preserve context across agent sessions and avoid forgetting user intentions.
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。
Do planreview step of the PLANNING phase of star-team workflow.
Do research step of the PREP phase of star-team workflow.
Run complex tasks successfully by utilitizing a development team with famour personalities. Use when
For research steps of the star-team workflow.
Do techspec step of the PLANNING phase of star-team workflow.
基于 SOC 职业分类
| name | report-keeping |
| description | Track task steps and write reports to preserve context across agent sessions and avoid forgetting user intentions. |
| allowed-tools | Bash, Read, Write |
Bureau helps you succeed by providing longer-term memory for (1) user requests and (2) plans, research and big-picture path taken.
LLMs suffer from drift (when mundanity overshadows the big things) and compaction (when details are simply lost). This is one of the leading causes of LLM coding failures requiring human interventions and killing productivity. We don't want that, do we?
Bureau assists in keeping a per-task directory with numbered report files. You create and read these files, Bureau only takes care of paths.
Run as scripts/bureau (relative to skill root) or just bureau (if installed globally).
Split your work into very clearly separated steps. Research, planning, reviewing plans, writing tests, writing implementation, reviewing code, updating docs, updating knowledge -- these should all be separate steps.
Before each step, re-read context from Bureau reports.
After each step, write a Bureau report file.
Run bureau without args to check current task and existing reports.
If current task doesn't match your work, run with -N <slug> to start a new task (or -S <name> to switch to an existing one IF user asked for that).
If user provided input, save it -- run with -n user-something (user-request, user-revision, user-feedback, etc.) and write their input VERBATIM, then append your interpretation.
Re-read recent reports to restore context.
Planning and review steps: read AS MANY reports as possible.
Execution steps: read all user requests + recent plans + all execution reports after the most recent plan.
Run with -n <suffix> to get next report filename (e.g., plan, impl, fix, research).
Write a report including:
