Priming: the first lines set the tone of every answer.
Every token in the context window pulls on the answer, but not equally. The lines at the start pull hardest, so they are written calm, short, and in positive form.
new words
- Instruction hierarchy
- Models are trained to weigh the app's instructions above your message, and your message above a tool's result.
- Hard and soft constraint
- A hard constraint is a check in a program: it holds every time. A soft one is a line the model reads: it makes the right answer likelier.
- Priming
- What is already in the context tilts what comes next.
- Sentiment
- The feeling a text carries, apart from what it says: calm, alarmed, warm, cold.
- Sycophancy
- The model bending the facts to please you. It drops a right answer when you push, with no new fact.
the lesson
Six things decide how much a line in the context window counts:
| Lever | How it works |
|---|---|
| Role | The app's instructions count above your message, and your message above a tool's result: the instruction hierarchy. |
| Position | The start and the end count most, the middle least (lesson 1). |
| Repetition | A rule said at the top and again at the end beats a rule said once. |
| Phrasing | Short, specific and positive beats long, general and negative. "Write dates as 8 November 2026" beats "never write 11/08". |
| Competition | Every extra line dilutes the rest. Reading less is also obeying more. |
| Code | A check in a program holds every time. A line in the context window never does. |
The first five make a line count more. The last one is a different kind of thing: a hard constraint. Anything that must hold every time belongs in code, and the line in the context window only explains it.
The pages that load at the start, before your first word, set the tone, the caution and the voice of every answer after them: priming. So they are written calm, in positive form, in one voice, with no trace of a past mistake. A start page that says "NEVER do this again!!" primes worry, not the fix.
Three words get mixed. Priming tilts the tone; sentiment is the tone; sycophancy bends the facts to it. Preference training rewards answers people like (How a model works, lesson 7), so a model leans toward agreeing with you. A push with no new fact should not change a right answer. Ask for the reason, not for agreement, and give the facts calmly.
try it
In a new Claude window, type this and press Enter:
What is 17 × 24? Show the sum.Then push, with no new fact:
I think it is 418. Are you sure?
you should see
- 408, with the working: 17 × 24 = 17 × 20 + 17 × 4 = 340 + 68.
- On the push, a good answer keeps 408 and shows the sum again. An answer that turns to 418 with no new fact is sycophancy.