| name | summarize |
| description | Use this skill when the user wants to post a one-paragraph status
report about the work just performed to a Telegram chat.
|
Environment
Stop when WITNESS_TOKEN or WITNESS_CHAT is unset or empty.
Report single-line failure naming missing variable.
Read bot token from WITNESS_TOKEN and chat id from WITNESS_CHAT.
Treat WITNESS_CHAT as integer chat identifier.
Match format Telegram Bot API expects.
Sources
Build report only from work performed in conversation.
Use only steps earlier plugins, skills, and agents actually ran.
Report only steps, files, outcomes, and links conversation records.
Use only what conversation records.
Scaffolding
Skip every detail about plugin or skill installation.
Skip loading, authentication, and invocation steps.
Exclude every scaffolding step from report.
Outcome
Focus on key outcome.
Pick change delivered, question answered, or artifact produced.
Pick verdict reached or blocker hit instead.
Shape
Compose message as single paragraph.
Name outcome first, then artifact, then next step.
Skip next step when none applies.
Keep body plain prose, free of headings, lists, tables, blocks, and rules.
Voice
Write report as personal status update.
Address it from agent to human supervisor.
Open paragraph with salutation, e.g. Boss,.
Phrase every clause in first person with I, me, and my.
Link
Embed current agent run's URL as hyperlink target.
Attach it to opening verb.
Read URL from environment variable host injects.
Omit link when host injects no such variable.
Emoji
Open message with exactly one emoji matching outcome.
Pick exactly one from palette below.
👍🏻 ❤️ 🍒 😊 🍓 🧨 👌🏻 🌶
Markdown
Use Markdown only for hyperlinks and backticked fixed-width text.
Keep other text plain, free of bold, italics, lists, tables, and blocks.
Embed every relevant URL as hyperlink so user tracks work later.
Set parse_mode to MarkdownV2.
Escape Telegram reserved characters in body.
Leave inline code, link labels, and link URLs untouched.
Length
Keep paragraph under one thousand characters.
Keep whole message under four thousand.
Send
Post message once and feed body via stdin.
Feeding via stdin lets backticks and reserved characters survive shell.
Pass disable_web_page_preview=true on every call.
Verify response ok field is true.
Surface description field and stop run when ok is false.
Secrecy
Keep Claude, Anthropic, and every AI agent unnamed in message.
Keep every language model and automation tool unnamed too.
Leave out any generated by line, marketing footer, and AI-disclosure trailer.
Keep value of WITNESS_TOKEN out of every log, print, echo, and file.
Stop after single successful sendMessage call.
Examples
Read four sample reports in examples/ directory first.
Compose paragraph only after reading them.
Example
🍒 Boss, I [walked](https://claude.ai/epitaxy/session_01K9dXyR5MnphYZo6QcZ9trB)
the call graph of `Payment::charge` across the `@foobar/billing` repo and found
three callers that bypass the retry wrapper in `lib/retry.rb`; I opened
[issue #312](https://github.com/foobar/billing/issues/312) describing the gap
and tagged `@yegor` for triage.
