for anything customers touch

Could you defend it?

Your agent made dozens of decisions you never saw. Some were choices. Many were just its habits. Whiteboard Defense tells you which is which, before someone asks.

“Why did you do X instead of Y?” is a fair question to get on a Tuesday.

curl -fsSL https://raw.githubusercontent.com/avestura/whiteboard-defense/main/install.sh | sh

irm https://raw.githubusercontent.com/avestura/whiteboard-defense/main/install.ps1 | iex

claude plugin marketplace add avestura/whiteboard-defense
claude plugin install whiteboard-defense@whiteboard-defense

curl -fsSL https://raw.githubusercontent.com/avestura/whiteboard-defense/main/install.sh | sh -s -- codex

curl -fsSL https://raw.githubusercontent.com/avestura/whiteboard-defense/main/install.sh | sh -s -- opencode

One command: installs for every agent it finds on your PATH.

What you get after a task

A short brief, and a full one saved to .whiteboard/. Every decision carries a tag saying who made it. The tag you care about is the red one.

Share-link service for invoices (example)

Decision	Choice	Who	Not taken
Link credential	Bearer token in URL	USER	Login-gated view
Token storage	SHA-256 hash	INFERRED	Plaintext
Expiry	7 days, fixed	GUESS	Per-link setting
Revocation	None	DEFAULT	Delete endpoint

Defend it: Where does this fail? A leaked link can’t be cut off before it expires, and every hit costs a full PDF render.

Paste next time: Every shareable link must be revocable by its creator.

nobody asked for this. the agent just never built it.

What it looks like in Claude Code

In Claude Code the plugin also draws in the interface: a band above the prompt, a brief pane, an entry in the status line, and briefs that start on their own once enough code has changed. These are working replicas built in HTML. Press the buttons.

claude · ~/billing-api

Opus 5.5 · ~/billing-api · main

Replayed from a recorded eval run. Switch to prompt to see the plain skill with no interface changes (/whiteboard-ui prompt).

the turn ends, the plugin counts the changed lines and asks for a brief by itself.

then try [ Open ] and the copy buttons in the pane.

claude · ~/billing-api · quiz

Opus 5.5 · ~/billing-api · main

Quiz mode works in every agent. Here the answers are scripted; in a real session the agent grades your own words against the code.

pick an answer. a wrong one gets corrected, not just marked.

Where agents lean on their defaults

Telling a model to “avoid a generic look” mostly swaps one default for another. Naming the specific pattern works much better. The catalogue ships with entries like these and grows through pull requests.

D-LOGO-01The monogram. Asked for a logo with no concept, agents reach for one letter, often in a circle, often with a dot. Avoid: a single-letter monogram, letters in circles, dot accents.
D-SLIDES-01The same deck. Click to advance, N for notes, a progress bar, an agenda slide. Fine, unless you never chose it.
D-COLOR-01Indigo to blue. The gradient that shows up when nobody specifies a palette.
D-ERR-01Catch, log, continue. A broad try/catch that returns a default and hides the failure from the caller.
D-SEC-02Homemade sessions. A token in localStorage, default hashing parameters, no stated threat model.
D-SCOPE-01Extras nobody asked for. A config layer, a settings page, a toggle. More surface to defend.

It is not only design. Architecture, naming, errors and security all have defaults.

Using it

automatic: A 15-line chat digest after each substantial task; it reads no extra files and saves nothing, to stay cheap. Turn it off with one sentence or "auto": "off".
/whiteboard-ui prompt: Claude Code only: turn the band, pane and status entry off and keep the plain skill. /whiteboard-ui rich brings them back.
/whiteboard: The full brief for the current work.
/whiteboard full 142: A brief for a PR, branch, or path that already exists.
/whiteboard defaults: Only the defaults, each with an Avoid: line to paste into your next prompt.
/whiteboard quiz: One question at a time, graded against the actual code. This is the part that makes you learn it.

Claude Code: /whiteboard-defense:whiteboard · Codex: $whiteboard · OpenCode: /whiteboard

PoCs, demos and experiments are exempt. Mark a project "tier": "poc" and it stays quiet.

What it can’t do

xIt can’t read its own weights. DEFAULT means “you didn’t specify it and it matches a habitual pattern”, not “proven bias”.

xThe “why” is reconstructed from the conversation and the diff. Usually right, sometimes wrong. Shaky ones are tagged GUESS.

xIt doesn’t replace reading the code that matters. It tells you where to look.

xIn Codex, OpenCode and Claude Code’s prompt mode, automatic briefs depend on the agent honouring the skill. If one is skipped, run it by hand. Claude Code’s rich UI triggers them from the plugin instead.