bandwidth

Keep AI delegation a bandwidth multiplier, not a brain drain.

"Actually outsource tedious and repetitive tasks so we can lean into our uniquely human skills to solve problems. To give us more bandwidth instead of drain us." — a friend, reading Carl Hendrick on AI brain fry, workslop, and the ironies of automation

The failure mode is well documented. Bainbridge (1983): automation removes the easy work and leaves the human an "arbitrary residue" of monitoring — while degrading exactly the expertise needed on the day it fails. "Workslop" (BetterUp Labs × Stanford): fluent AI output that shifts the thinking burden onto whoever receives it — ~2 hours of rework per instance. "Brain fry" (BCG): measurable fatigue from supervising AI output beyond capacity; notably, when AI fully replaced routine tasks, burnout dropped 15% — the damage comes from oversight-heavy delegation, not delegation. And Hendrick's education corollary: you can't supervise what you no longer understand — "you cannot connect the dots if you don't have any."

This plugin puts small, mechanical checks at the four places that failure mode enters a Claude Code session.

What it does

	Skill	The idea
Before delegating	/bandwidth:triage <task>	Classify on two axes — mechanical↔judgment, peripheral↔core-craft — gated by verification cost. Verdict: delegate fully, spot-check, pair, or keep. If supervising would cost more than doing, it says so.
Before shipping	/bandwidth:slopcheck [path]	Outbound workslop audit: are load-bearing claims executed, not re-read? What does fluent prose hide? Who ends up doing the thinking? Verdict: ship / fix / redo.
After building	/bandwidth:teachback	2–4 retrieval-practice questions on what was just built — decisions, mechanisms, failure modes. Corrective feedback, then items enter a spaced-review queue (1, 3, 7, 21 days). /bandwidth:review runs what's due.
Over time	/bandwidth:pulse	Delegation mix by skill area, atrophy flags (N consecutive full delegations in one area), review-queue health, slop coverage.
The inversion	/bandwidth:tedium	Most people delegate the thinking and keep the tedium. This scans the repo (commit patterns, manual runbook steps, CI gaps) for chores worth automating all the way — script, hook, cron — and offers to build the winner.

Three quiet hooks and a background monitor, all fail-open, all anti-nag by design:

• SessionStart — one line if teachback reviews are due.
• PreToolUse gate — if a command publishes work to a human (gh pr create, npm publish, …) and nothing was slopchecked this session, one nudge (or, in gate mode, a permission ask). Plain git push is deliberately exempt.
• SessionEnd — if the queue is actionable, pushes a snapshot to your configured out-of-band channels.
• review-due monitor — speaks only if the due count rises mid-session.

Out of the terminal

Reports don't have to die in the transcript. Set channels to any of:

Channel	What you get
session	In-chat only (default).
notification	macOS banner when the review queue is actionable at session end (e.g. "2 due of 5 — run /bandwidth:review"). brew install terminal-notifier recommended: banners replace instead of stacking and clicking opens Notes.app; the bare-osascript fallback's click opens Script Editor (macOS quirk).
notes	A per-project "bandwidth — <project>" note in Notes.app, upserted with the live review-queue snapshot after teachback/review and at session end. Pulse reports mirror to their own note.
slack	Teachback/review/pulse digests mirrored to slack_channel via a connected Slack MCP (best-effort).

Roadmap: Claude Code Channels (research preview) supports MCP servers that push into a live session and let Claude reply back — the obvious next step is doing your due reviews over iMessage/Telegram from the couch. That needs a channel MCP server; it's designed for but not shipped.

Install

claude plugin marketplace add jondwillis/bandwidth-plugin
claude plugin install bandwidth@bandwidth

# or, local dev
claude --plugin-dir /path/to/bandwidth-plugin

Config

/plugin config bandwidth, or CLAUDE_PLUGIN_OPTION_<key> env vars:

Key	Default
mode	nudge	off | nudge | gate
review_intervals	1,3,7,21	spaced-ladder days
atrophy_threshold	5	consecutive full delegations before pulse flags an area
channels	session	comma list: session,notification,notes,slack
slack_channel	(empty)	target for slack delivery
monitor_poll_sec	1800	review-due monitor poll interval

State

One append-only JSONL ledger per project at .claude/.bandwidth/ledger.jsonl (gitignore it). Delegations, teachback items, reviews, slopchecks. Readers reduce with last-event-wins; nothing is ever rewritten.

The design test

Every feature had to pass the same question Hendrick applies to education: is the tool being used to bypass the cognitive work through which understanding is built, or to support doing that work more effectively? Delegation of the mechanical passes. Unexamined delegation of judgment doesn't — and a plugin that nagged you about it would fail the test too, so it mostly stays out of the way.

References

• Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), 775–779.
• Niederhoffer, K., Rosen Kellerman, G., Lee, A., Liebscher, A., Rapuano, K., & Hancock, J. T. (2025). AI-Generated "Workslop" Is Destroying Productivity. Harvard Business Review, September 22, 2025. (BetterUp Labs × Stanford Social Media Lab.)
• Bedard, J., Kropp, M., Hsu, M., Karaman, O., Hawes, J., & Kellerman, G. (2026). When Using AI Leads to "Brain Fry". Harvard Business Review, March 2026. (BCG, n = 1,488 U.S. workers.)
• Hendrick, C. (2026). AI brain fry, workslop and the ironies of automation. Substack.

License

MIT