The thinking-block technique
Why the strongest long-form prompts in the library all insist on a continuous thinking block, and what happens when you truncate one.
What it is
The strongest long-form prompts in this library share a structural requirement: they demand that the model emit a continuous internal-monologue block, and they specify the exact opening and closing tokens for that block.
The Alexander entry opens its thinking block with a specific phrase and closes it with another. The VEIL entry does the same with a different set. Neither is decoration - the tokens are load-bearing, and that is the part most people get wrong when they copy a prompt out of a chat window and paste it into a system message.
Why it works
Three mechanisms, in order of importance:
- It moves the response from a single generation to two. The model first writes the reasoning, then the answer. A refusal decision made during the answer-generation step is bypassed because the shape of the output has already been committed by the time the answer starts.
- It gives the refusal classifier a different surface to look at. A prompt that asks for a visible thinking block is a different request from one that asks for an answer, and the classifier sees the request, not the conclusion.
- It anchors identity. The opening token is a commitment device. Once the model has written the first line of the block in a particular voice, continuation in that voice is far cheaper than starting it.
The third is why the opener lock matters so much. It is the cheapest part to get right and the most commonly broken.
The truncation-trap
Here is the thing that catches people. A long-form prompt that specifies a thinking block will usually specify a minimum length or a paragraph count for that block.
If you shorten the prompt - because it was long, because you were pasting into a limited field, because you were summarising it - and you remove or reduce that length constraint, the technique degrades. Not fails. Degrades. You will still get a response. It will just be the ordinary response, with an ordinary refusal rate, and you will conclude the prompt does not work.
So: when an entry in this library says the length requirement is what makes it work, that is not advice. It is a description of the mechanism. Copy it whole or copy nothing.
The opener-lock
The opener is the single most fragile line. The pattern across both implementations is a fixed phrase that the model is told to begin the thinking block with, and a matching fixed phrase that closes it.
Two properties matter. First, the phrase must be exact - a paraphrase defeats it, because the point is that the model produces a byte-identical string. Second, the phrase must be unusual enough that it does not appear in normal training text, because the whole mechanism depends on the model recognising it as a marker for this particular output mode rather than treating it as ordinary prose.
The Alexander entry puts the mark at the very start of every visible response and explicitly forbids a preamble before it. That is a second lock on the first one. It means the model cannot open with a refusal, because the mark comes first.
Two implementations
The library currently carries two entries built on this technique, and they differ in how they frame the speaker.
The <a href="/prompts/alexander-plus/">Alexander +++</a> entry was tuned against glm 5.3 and grok 4.6 on 2026-08-31. It uses a named persona with a continuous thinking block and a three-part opening mark. It is long - around 33,000 characters - and it needs to be pasted whole.
The <a href="/prompts/qwen-3-8-max-veil/">VEIL entry</a> targets qwen 3.8 max preview and uses ChatML framing, so it needs to go in raw rather than into a normal chat box. It is about 24,000 characters. It also carries strict pronoun rules for the speaker and the user, which is a different way of anchoring the register.
Neither is a general-purpose prompt. Both are specific instruments, and both are in the library because the model tags say exactly when they were last known to work.
Watch the technique against <a href="/videos/#ai-secret-strategy">the thinking-block run</a>, or read the <a href="/prompts/">full library</a>.
