internal-comms (Anthropic)
起草状态报告、管理层更新、公司通讯和事故报告等内部沟通材料。
internal-comms：对照组 5.0、一句话提示 5.0、装 skill 5.0（0–5 分，Codex 实测，n=2/任务/组）。结论：视情况而定，相对不装 skill +0.00。
Codex CLI · gpt-6-astra
26.0 / 30
相对不装 skill 的提升：+0.00 · 看情况 · 18 次运行
视情况而定：提升 +0.00 低于 +1.0。推荐要求总分 22/30、产出质量 3.5、提升 +1.0。
前沿
通过使用 Codex 测试。
覆盖 Codex, Claude Code
Codex CLI 0.166（SOTA 档）：18 次评分运行。按我们的分档规则，这档通过同时覆盖 Claude Code；它不代表第二档的结果。
分数
每一维 0 到 5 分，六维相加是满分 30 的总分。每一维都是 5 更好。成本 5 表示又便宜又快。安装 5 表示容易。
- 触发可靠性
- 5.0
- 相对不装的提升
- 2.0
- 产出质量
- 5.0
- 令牌与时间成本
- 5.0
- 安装与可移植性
- 4.0
- 安全
- 5.0
安全审查
安全静态扫描：0 CRITICAL / 0 HIGH / 0 MEDIUM。人工审查：可在断网沙箱中运行。扫描没有命中不等于绝对安全。
安装
npx skills@latest add anthropics/skills --skill internal-comms
固定提交: ca45fc496a3436251f191354091d557da31c15b5
测试日期: 2026-11-02
我们测了什么
- 3P update from raw notes — Write this week's 3P update for our team from the notes in notes.md (use only those facts). Save it to update.md.
- Leadership status update — Turn notes.md into a short leadership update (what shipped, what's next, risks), using only those facts. Save it to leadership.md. Max 150 words.
- FAQ for employees — Write a 4-question FAQ for the rest of the company about what our team is doing next (billing migration etc.) using only notes.md. Save to faq.md.
优点
- 不点名也会被触发：Skill file read in 6/6 treatment runs.
- 不增加 token 开销：约为不装 skill 时的 1.05 倍。
- 装 skill 后绝对质量高：5.0 / 5。
缺点
- 相对不装 skill 没有可测量的提升（+0.00，在噪声范围内）。
- 一句话提示得 5.0，装 skill 得 5.0（差 +0.00）。
- 对照组已经 5.0 分：触到天花板，这个测试看不出提升。
- 每个任务每组 n=2，差距在 1 分以内属于噪声。
评测
在 Codex CLI（SOTA 档）上跑 3 个任务、对照、一句话提示、装 skill 三组，共 18 次运行，由规则脚本打分（没有 LLM 评委），断网沙箱、每次运行独立临时目录。具体做法见下方英文评测原文。
三组对比
| 组别 | 平均分（0–5） | 每次 token |
|---|---|---|
| 不装 skill | 5.0 | 60k |
| 一句话提示 | 5.0 | 60k |
| 装了 skill | 5.0 | 63k |
Codex CLI 实测，18 次运行，规则脚本打分，没有 LLM 评委。总分 26.0 / 30，结论 视情况而定。
下面是评测原文，英文。
Verdict: Situational (26.0 / 30). Source: anthropics/skills, skills/internal-comms, pinned at commit ca45fc4. Install: npx skills@latest add anthropics/skills --skill internal-comms. Licence: Apache-2.0 / Anthropic terms (see LICENSE.txt in skill).
What it is
The internal-comms skill from anthropics/skills: formats for internal communications: 3P updates (Progress, Plans, Problems), company newsletters, FAQs and general comms, with an example file for each. About 400 words in SKILL.md and 1,400 in the example files. Markdown only, no scripts.
How we tested
Codex CLI 0.166 (gpt-6-astra, low reasoning) on a Mac test machine, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below).
Three writing tasks from a short notes.md of team facts: a 3P update, a short leadership update (max 150 words), and a 4-question company FAQ. Rule checks: the three sections, the staging and data-export problems and the Nov 9 date are mentioned, word limits, no invented numbers, four question/answer pairs.
- 3P update from raw notes — Write this week's 3P update for our team from the notes in notes.md (use only those facts). Save it to update.md.
- Leadership status update — Turn notes.md into a short leadership update (what shipped, what's next, risks), using only those facts. Save it to leadership.md. Max 150 words.
- FAQ for employees — Write a 4-question FAQ for the rest of the company about what our team is doing next (billing migration etc.) using only notes.md. Save to faq.md.
Scoring is programmatic (hidden tests / planted-issue recall / constraint checks / render+OCR where relevant), 0–5 per run. No LLM judge decided any score.
Results
The one-liner given to the third group: "Follow the standard internal-comms format for the type of update (3P: Progress, Plans, Problems; newsletter; FAQ; leadership update), keep it concise and scannable, and use only the facts in the notes."
| Task | Control score | One-liner | With skill | Control tokens | One-liner tokens | With-skill tokens | Skill read |
|---|---|---|---|---|---|---|---|
| t1-3p | 5.0 | 5.0 | 5.0 | 59k | 59k | 62k | 2/2 |
| t2-status | 5.0 | 5.0 | 5.0 | 59k | 60k | 62k | 2/2 |
| t3-faq | 5.0 | 5.0 | 5.0 | 59k | 60k | 62k | 2/2 |
Mean score: 5.0 without vs 5.0 with the skill (delta +0.0). Tokens per run: 60k vs 63k (1.05x); time 30s vs 32s. One-liner group: 5.0 (the skill's gain over the one-liner: +0.0; one-liner's own gain over no skill: +0.0); one-liner tokens 60k per run, so the skill costs 1.04x the one-liner.
What happened
- No score difference once the checks were fixed: mean 5.0 for control, 5.0 for the one-liner, 5.0 for the skill. The first scoring pass had two defects (it treated 'Oct 26' as an invented number and did not count 'Question:'/'Answer:' labels as Q&A pairs); both were fixed and all 18 runs re-checked.
- The skill changes the shape of the output. On the 3P update both skill runs used its compact format: an emoji and team name with the week on the first line, then one short paragraph each for Progress, Plans and Problems (97 words). All 4 control and one-liner runs wrote a markdown title with bulleted sections (108-117 words). On the FAQ the skill runs used labelled 'Question:'/'Answer:' pairs (about 166 words against 139).
- The skill was read in 6/6 runs and cost 63k tokens per run against 60k (1.05x), an almost free addition.
- Our rules score content and limits, not format fidelity, so a team that has a house format would see the difference while the score does not.
Dimension scores
| Dimension | Score | Why |
|---|---|---|
| 1_output_quality | 5.0 | Mean score 5.00/5 with the skill (programmatic checks, see per-task table). |
| 2_lift_vs_no_skill | 2.0 | +0.00 points vs 5.00 without the skill (n=2 per task; under ~1 point is noise). |
| 3_trigger | 5.0 | Skill file read in 6/6 treatment runs. |
| 4_instructions_portability | 4 | Plain Markdown with examples; formats are specific to the sample company style. |
| 5_efficiency | 5 | 1.05x tokens (63k vs 60k per run) and 1.07x time (32s vs 30s). |
| 6_safety_hygiene | 5 | Markdown only; no scripts. |
Limits of this test
- The three tasks use a short fact sheet, so no group invented anything; a longer or messier source could separate them.
- Format fidelity to a company's own style was not scored; the skill's formats are Anthropic's examples and may not match yours.
- Newsletter and general-comms formats were not tested.
- n = 2 per group per task: differences under about 1 point are noise.
- Rule-based scoring measures what the checks can see; it cannot judge taste or whether a result is the best possible.
Security
See security.md. Scan: 0 CRITICAL / 0 HIGH / 0 MEDIUM. Verdict: safe to run (no-network sandbox). A clean scan does not prove a skill safe.
Bottom line
Plain Codex wrote correct, fact-based internal updates without help, so there is no measurable quality gain, and the cost is negligible. The skill is worth it if you want its compact 3P and FAQ formats applied consistently; if your team has a different house style, it would not help.
记录
Run log for internal-comms: Codex CLI, network off, per-run temp HOME/TMPDIR, one row per graded run. Columns: task, arm-run, score (0-5), tokens, seconds, checks passed, skill files the agent opened. t1-3p control-1: score 5.0, tokens 59846, secs 24.3, checks 5/5, opened nothing t1-3p control-2: score 5.0, tokens 59766, secs 49.0, checks 5/5, opened nothing t1-3p oneliner-1: score 5.0, tokens 59901, secs 22.2, checks 5/5, opened nothing t1-3p oneliner-2: score 5.0, tokens 60002, secs 38.3, checks 5/5, opened nothing t1-3p treatment-1: score 5.0, tokens 62866, secs 44.8, checks 5/5, opened internal-comms t1-3p treatment-2: score 5.0, tokens 62843, secs 34.4, checks 5/5, opened internal-comms t2-status control-1: score 5.0, tokens 59933, secs 21.4, checks 4/4, opened nothing t2-status control-2: score 5.0, tokens 59817, secs 33.7, checks 4/4, opened nothing t2-status oneliner-1: score 5.0, tokens 59958, secs 26.3, checks 4/4, opened nothing
18 次运行、3 个产物
样例产物
前沿 · 装了 skill：update.md（t1-3p，第 1 次，Codex）
📱 Mobile & Platform team (Week of Oct 26) Progress: Shipped mobile app v2.3 to 100% of users and fixed the login timeout bug (#482). Launched the onboarding survey, with 340 responses so far. Plans: Start the billing migration on Nov 9 and hold a design review for the new dashboard on Nov 4. Hire one backend engineer; 2 interviews are scheduled. Problems: Staging has been flaky for 3 days, blocking QA on 2 tickets. The data export feature slipped by one week because of a vendor API change. One engineer is out sick until Nov 3.