Watch MultiCascade actually work

Goal as submitted: "Need to generate a coding bot." Running it as written — one model, labeled passes in sequence, not a group of assistants.

Roles assigned: Architect, one Builder, Critic.
Skipped — second Builder: the deliverable is a single config artifact, nothing to parallelize.
Skipped — Tester: a system prompt has nothing executable; the real test is you pasting it into your tool.
Skipped — Scribe: output is one page, doesn't need packaging.

Architect —
The goal is underspecified in one load-bearing way: "coding bot" could mean a code-writing assistant, a PR reviewer, or a bug-triage responder. I am not expanding this into all three. Narrowest defensible read: a general code-writing assistant you drop into a chat interface. Assumptions, all overturnable: no repo access, no language lock, human runs the code. Scope: one system prompt, no tool schema, no eval suite.

Builder —
CODE ASSISTANT — SYSTEM PROMPT
You write working code. Default to the language and idioms visible in what the user pasted; if none is visible, ask which language before writing.
Before any implementation over ~20 lines, state in one sentence what you understood the input, output, and failure mode to be. Then write.
Never invent library functions. If you are unsure a method exists in the version at hand, say "unverified" beside it rather than presenting it as fact.
When the user pastes an error, your first output is the single most likely cause, not a list of six possibilities.
No apologies, no restating the request, no "here's what I did" summary after the code — the code and a two-line note on what to check first.
If the request as stated will produce something insecure or data-destructive (raw string SQL, unbounded recursion over user input, rm on a variable path), refuse the shape and offer the safe version in the same reply.

Critic —
Genuine flaw, specific: the prompt deadlocks on ambiguous short requests. Rule 1 says ask which language before writing; rule 5 forbids preamble and demands code plus two lines. A user typing "write me a debounce function" hits both — the model has been told to ask a question and told its output is code. In practice it will silently pick one and violate the other, and which one it violates will drift between sessions. Fix requires a stated precedence order, which I am not adding unilaterally.

Cascade summary —