See llms.txt for all machine-readable content.
Toolbox talks get skipped when the foreman has to write one before the shift starts, and nobody can prove later that they were held. This workflow picks today's topic from an Airtable library without repeating recent ones, has a DeepSeek model write the talk, checks it, has ElevenLabs voice it, and sends it to the foreman on Telegram. The foreman replies done to sign it off, and a Friday email shows what was delivered and what was confirmed.
Last updated: November 2026.
Every morning at 05:30 the Toolbox Topics table and the last two weeks of the Toolbox Log are read. Today's topic rotates over active topics that were not used in the last 14 days, and an empty library sends a Telegram alert instead of failing quietly. The model writes a 130 to 170 word talk, and a quality check flags talks that are too short or too long, skip the topic or quote a percentage that was not in the source. Good talks are voiced by ElevenLabs and sent as an audio note. Flagged talks, and any talk where the voice step fails, go out as text with the reason. Every delivery is logged with its type and Telegram message id, and the topic gets its last_used date. When the foreman replies done, optionally with a crew count, the log row is marked held. A 10:00 check nudges the foreman if today's talk is still unconfirmed, and the Friday summary reports delivered, confirmed and text-only talks.
How does the foreman sign off a talk?
By replying done in the Telegram chat, or done 12 to record the crew count.
Can the same topic come back two days running?
No. Topics used in the last 14 days are skipped unless the library has nothing else.
What if the AI writes something off?
Check Talk Quality sends a text version with the reason so the foreman reads it first.
What if nobody confirms?
At 10:00 the foreman gets a nudge, and Friday's summary lists the talk as not confirmed.
Built with n8n. Need an assessment on your business? Feel free to reach out at https://khmuhtadin.com/consultation/