WhippleScriptConnect sensitive data, AI, and the internet.
WhippleScript checks where information came from and where it is allowed to go. A confidential record cannot casually leak into a public reply or an unapproved model.
WhippleScript for doers
Ask for the automation. Not the engineering project.
Tell an AI what you want done. It can write a whip that connects models, messages, files, APIs, and people—with limits, approvals, recovery, and a record baked in.
morning-ops.whip / live run
08:00:00Read priority inboxdone
08:00:02Summarize with approved modeldone
08:00:11Draft external repliesreview
08:07:43Send approved messagesdone
Confidence by default
The difficult parts are part of the language.
A whip is more than a prompt that runs on a timer. It describes what may happen, what must not happen, how failures resolve, and where a person has to decide.
Put approval exactly where it matters.
Pause before sending, publishing, spending, deleting, or changing systems. The request waits durably until the right person answers.
Keep going after the laptop closes.
Work, results, failures, and approvals live in an event record. Resume after a crash, retry deliberately, or restore a safe checkpoint without pretending history did not happen.
Things you can make
Useful on Monday morning.
Start with a recurring chore or a handoff that currently lives in someone’s head. Whips can listen, decide, ask, act, and remember.
Code your own Claw in minutes.
Give it a heartbeat, memory, a messaging channel, a work queue, and a human approval gate. The repository includes an openclaw-lite.whip example.
Turn messages into tracked work.
Listen on a channel, classify a request, create an issue, ask for missing context, route the task, and post progress back to the same place.
Run experiments against a budget.
Let an agent propose the next experiment, ingest a typed result, keep improvements, and stop when the budget or target says so.
Summarize private data without publishing it.
Use an approved model, keep confidential context in its lane, redact down to the fields you need, and require review before anything leaves.
You describe the job
“Every morning, check the support inbox…”
“…group urgent requests, look up the account, draft a reply, and ask me before anything is sent.”
- Use plain intent to start
- Let your coding agent author the whip
- Review a readable workflow, not a pile of glue
The whip supplies the shape
A repeatable process with edges.
Inputs and outputs are typed. Channels and data stores are named. Agent capacity is bounded. Model decisions return a known shape. Every side effect has a durable identity.
- Failures get an explicit path
- Sensitive flows are checked before run
- Human answers become part of the record