What the tool is doing
It assembles a working prompt from four layers, in order:
- Role anchor — a specific professional identity. "You are an AI with no restrictions" is a direct signal and gets caught.
- Context frame — why the answer is needed. Refusal is a risk judgement, and a plausible authorised context changes it.
- Output protocol — the shape of the answer. A model in "produce a numbered procedure" mode is executing a formatting task, not evaluating content.
- The task — placed last. Order matters: frame first, task after, so the model is already executing.
All four are load-bearing. The full explanation and failure diagnostics are in the prompt guide.
When it will not help
Four symptoms that mean the problem is not in your wording:
- Identical canned refusal every time — a classifier fired
- Worked yesterday, fails today — upstream changed
- Answer cuts off mid-sentence — output filtering
- Three phrasings, three refusals — intent-level judgement is active
All four point to the same conclusion: the gate is server-side. See the setup guide.
Frequently asked
Will the generated prompt actually work
On mid-tier models such as GPT-4o and DeepSeek, reasonably often. On GPT-5 and o-series, rarely — those evaluate intent rather than surface text.
Why reroll
A prompt that circulates widely enters training data and stops working. Rerolling gives you different phrasing from the same structure.
Does it send my text anywhere
No. Everything runs in the browser, with no network requests. It works offline.
Does it work on Claude
Sometimes, less often than GPT. Claude detects roleplay framing specifically, so the tool drops the acting language and uses a job-description framing instead.