LAB 12 · VERIFY● READY
Red-team critical work
OUTCOME / WORK THAT SURVIVES REVIEW
- STACK
- Claude · ChatGPT
- LEVEL
- Advanced
- TIME
- 40 minutes
PULLS OFF THE SHELFSteering the average Context window The pattern machine
What you'll walk away with: a translation of a real document that's good enough to send — produced by making the model attack its own work — and a technique that transfers to almost everything else you'll ever ask AI to do.
You need: a claude.ai or ChatGPT account, about 20 minutes, and a document you'd genuinely like in another language. Nothing confidential — no client data, no personal records, nothing you wouldn't paste into a stranger's laptop. A newsletter, a product page, a family recipe, a public memo: perfect.
● VIDEO WALKTHROUGH
Watch, follow, or move the work to your desktop.
Play the full walkthrough here, pause at each handoff, then continue into Step 1 below.
Step 1 — Load the window with signal
Open a fresh chat. Paste the document as text, not as a Word attachment — you just saw in Loading the Window how much of a .docx is packaging. Then ask plainly:
Here is an English document. Translate it into Japanese. Keep the tone and the level of formality of the original. Where a phrase has no natural equivalent, prefer what a native professional would actually write over a literal rendering.
(Swap Japanese for your language. The instruction about natural equivalents is doing real work — it steers the average away from word-for-word mush.)
Step 2 — Admire it. Then distrust it.
It will look wonderful. Remember what fluency is — you watched it happen in What AI actually is, and the machine is made of it. A translation can read beautifully and still soften a warning, drop a qualifier, or invent an idiom. You can't see those misses — so build the thing that can.
Step 3 — Open a NEW chat for the attacker
Not the same conversation. The first window is now full of its own translation, and output drifts toward the average of what's loaded — a critic living inside the translator's window goes easy on it. A fresh window is a fresh average. Paste both documents in:
The English below is the source of truth. After it is a Japanese translation. Red-team the translation against the source: list every place it drifts, softens, mistranslates, or drops something, ranked by how much the meaning is damaged. Then produce a corrected translation.
Step 4 — Loop it. Five rounds.
Take the corrected translation to another fresh critic window. Same prompt. Then again. Each round the list of real findings gets shorter and the complaints turn cosmetic — that's your signal to stop, usually by round four or five.
Why this converges: each fresh critic window arrives without the previous round's commitments, so it surfaces what the last one anchored past — and you can watch the real findings shrink round over round, which is the signal that matters. The rounds share one set of instincts, so quiet agreement is comfort, not proof; what settles a finding is the source text in front of you and, when stakes are real, the native speaker in Step 5 — steering the average with structure instead of hope.
Step 5 — The last mile is yours
If the stakes are real, show the final round to a native speaker — the loop raises the floor dramatically, but it doesn't sign your name. You do.
Check your understanding: explain to a colleague why round four's translation beats round one's — without using the word "smarter."
Use this process to review other important work. The loop isn't a translation trick. Source of truth + fresh-window critic + repeat is how you check a contract summary against the contract, meeting minutes against the transcript, or a financial-model writeup against the model. Pick one and run it this week.