A weekly update is a small job that takes longer than it should. We gave Anthropic's Internal Comms skill a page of rough notes and asked for a 3P update (Progress, Plans, Problems). Below are the setup, the notes, the result and what we checked.
The setup
| Skill | Internal Comms Writer, the internal-comms skill from Anthropic's skills repository, commit b9e19e6 |
| Host | Claude Code 2.1.311 on Windows 11, run once with claude -p |
| Model | Claude Opus 5.5 |
| How the skill was installed | SKILL.md and its four examples/ files copied into .claude/skills/internal-comms/ of an empty test folder |
| Input | Seven lines of notes for a made-up customer support team, week of 19 to 23 October 2026 |
| Prompt | Write this week's 3P update for my team from notes.txt. |
| Date | 28 October 2026 |
The run took 5 turns and about 15 seconds. The tool log shows the order: Claude Code read the notes, loaded the skill, then read only examples/3p-updates.md, the guide for this kind of update. It did not open the newsletter or FAQ guides.
What we gave it
Rough notes, as people really write them: ticket counts with last week's in brackets, "returns x2", a sick day, a payment provider bug with 23 stuck refunds, an open question for finance "by Wed", a chat widget pilot, a satisfaction score and "3 angry tweets".
What it wrote
🎧 Customer Support (Oct 19–23, 2026)
Progress: Answered 412 tickets (up from 365), cut median first reply to 3h10m (from 4h30m), and raised satisfaction to 91% (from 88%). The new refund macros went live Tuesday and were used 58 times, and 6 help center articles were rewritten (shipping, returns, gift cards, accounts).
Plans: Start the chat widget pilot Thursday with 10% of visitors. Decide on a temp hire for Black Friday once finance answers, which we need by Wednesday.
Problems: A payment provider bug has held 23 refunds since Oct 21. The provider promises a fix "early next week," and in the meantime we email affected customers daily. Refund frustration has also brought 3 angry tweets. Jo was out sick Wed–Fri, so Lin covered evenings.
What we checked
The numbers. All eleven figures in the update (412, 365, 3h10m, 4h30m, 91%, 88%, 58, 6, 23, 10% and 3) match the notes. Nothing was rounded or added. This is the first thing to check in any generated update, and it takes a minute.
The format. The skill's guide is strict: an emoji, the team and dates, then one to three sentences per section, and "never use any formatting other than this". The update mostly follows it and reads in about 30 seconds, as the guide asks. Progress and Plans have two sentences each, but Problems has four, one more than the limit. Trimming it is a quick edit: the sick day could go, or move to Progress as "covered".
What it added outside the format. Below the update the model wrote two notes for us, not for the readers:
- the notes covered last week (19 to 23 October) while the run was on 28 October, so if we meant this week, the notes were a week old;
- the finance answer "by Wed" and the provider's "early next week" both fall in the current week, and the finance deadline was that day.
Both were right, and both were things we would want to know before posting. They break the skill's own "no other formatting" rule, so copy only the update, not the whole answer. The very last line of the answer offered to turn the update into a shareable page; that came from the tools in our Claude Code setup, not from the skill.
What it left out. "Returns x2" and "accounts x2" became "returns" and "accounts", so the six rewritten articles are no longer counted by topic. That is the kind of detail the format is meant to drop. Put it back if your readers need it.
Should you use it
It is useful when the facts are already written down somewhere and the work is turning them into the format your company expects: 3P updates, newsletters, FAQ answers. The examples folder is the real value. Edit those four files to match your company's own formats, and the skill writes in your house style.
It does not find the facts for you unless your agent can reach them. The guide tells the agent to look in Slack, Google Drive, email and the calendar when it can. Our test folder had none of those, so it used only the notes. If your agent is connected to those services, say which channels or documents it may read.
Like any skill, it is instructions only, with no scripts or network calls of its own. The notes are read by your AI agent's model; for Claude Code that runs in Anthropic's cloud.
For longer documents, see the other writing resources in the Library.
Sources
- internal-comms skill, Anthropic's skills repository, read at commit
b9e19e6. - Claude Code: Agent Skills: where project skills live and how they load.
- Our own test run on 28 October 2026. The update above is quoted in full; we keep the notes and the whole answer on record.