Learning Dashboard · Reading & Auditing AI-Generated Code

Help

Workflow Guide

The Active Iteration Loop

This dashboard runs beside Claude Code in a tight loop: read a milestone, hand its rules to the agent, let it plan and implement, then prove the result with the drill. One concept per loop keeps diffs small and reviewable.

Pick a milestone tab (1–5) and read the left Foundational panel — the three modules and the clickable code anatomy. Click any underlined token to see what it does at runtime.
Switch Claude Code into Plan Mode: press Shift+Tab to cycle modes until the prompt shows plan mode on. In Plan Mode the agent proposes a plan and changes nothing until you approve.
Open the right Auditing panel and click Copy CLAUDE.md Rules. Paste them into your project’s CLAUDE.md so the agent follows that milestone’s conventions on every turn.
Describe the change, paste the milestone’s focus, and let Claude return a plan. Read the plan against the milestone’s rules before approving it.
Approve, let it implement, then review the diff. Keep the change scoped to one concept and reject drive-by refactors.
Run the drill: open the bottom Assessment panel, click Copy Adversarial Prompt, and paste it into a fresh LLM chat to test your own understanding.
Commit with a clear, scoped message — stage selectively, then commit: git add -p git commit -m "M2: validate Product contract at the boundary"

The Fail-Safe Commands

Every Copy Adversarial Prompt ends with a fail-safe protocol. When you are running a drill in an LLM chat and get stuck, type one of these bracketed commands and the AI is instructed to honour it strictly. It will not mention them up front — invoke them only when you need them.

[HINT]

A subtle hint that narrows you to the right file section or concept — without revealing the bug. Reach for this first; it unsticks you while keeping the puzzle intact.

[EXPLAIN]

A line-by-line execution trace in plain English — what each line does as the code runs. Use it when you cannot follow the control flow or the data path.

[REVEAL]

Ends the drill immediately: the AI outputs the exact line number of the flaw, the corrected code, and an explanation of the architectural fix. Use it only when you want the answer.

Escalate gently: [HINT] → [EXPLAIN] → [REVEAL]. The more you solve before revealing, the more the drill teaches.

Environment Fallbacks

When the testing AI’s claims confuse you — or it insists something “works” — stop and run a hard, deterministic check in your terminal. The compiler and the test runner do not bluff: a green result is proof; a red one names the exact problem.

Type-check the whole project (no build)

npx tsc --noEmit

Compiles without writing any files and prints every type error with its file and line. Add --strict if your tsconfig is not already strict. No output means the types are sound.

Run the test suite once

npx vitest run # Vitest npx jest # Jest npm test # whatever the project wired up

Runs every test a single time and reports pass/fail counts. A red run tells you which behaviour broke; an AI’s reassurance does not.

Narrow to one file, or watch as you edit

npx vitest run src/chargeOrder.test.ts # one file npx vitest # watch mode, re-runs on save

Focus on the file you are auditing, or leave watch mode running so every save re-runs instantly.

Rule of thumb: a green tsc plus green tests beats any AI’s confidence. If the agent says “this works” but the terminal is red, trust the terminal.