Frame AI before you produce anything
An AI that hallucinates is an AI that never received a tight enough frame.
Two names, two things.
The Method is the discipline: clarify before producing, set safeguards, refuse to invent.
The Template is its tool: five blocks, in this order.
A model does not flag what it is missing: it fills the gap
The starting point is not an opinion about AI — it is how the tool actually works. Google sums it up in one line in its prompt engineering white paper: "Remember how an LLM works; it’s a prediction engine."
A machine that predicts does not stop at a gap. It predicts what belongs in the gap. The activity rate, the start date, the legal basis, the name of the labor agreement: everything you did not say gets written anyway, in a perfectly plausible form. And plausible content does not stand out on review, precisely because it is plausible.
The same document draws the consequence: insufficient instructions lead to ambiguous or inaccurate answers, and framing a request is iterative work, not a stroke of luck. We draw our own: this is not a flaw to fix, it is a property to design around.
"You do not operate an engine without understanding how it works. You do not operate an AI without understanding its architecture."
Source: Lee Boonstra, Prompt Engineering, Google, September 2024, p. 6–7.
You do not produce first: you clarify first
Common practice fires off the request, then corrects it ten times over. The Method reverses the order: it asks the missing questions before the first line is written. Three rules govern this phase, and all three are rules of restraint.
One question at a time
A fifteen-line questionnaire does not get an answer; it gets abandoned. One numbered question gets its answer.
Only what changes the deliverable
If the answer changes nothing about the outcome, the question does not get asked. Framing that questions for form’s sake gets resented, then bypassed.
What is missing gets written down
Unavailable information stays unavailable. It carries a visible marker in the rendered text — never a plausible-sounding value that would slip past review unnoticed.
These answers are not wasted: they fill the Data block of the Template. That is the whole relationship between the Method’s two halves — clarification is not a polite preliminary, it is how the material gets made.
The Palambur Template: five blocks, in this order
One control rule, and only one: a missing block means an approximate result.
Instruction
What to do, and why
A clean action sentence. The "why" is not decorative: it is what lets the model rule on cases you did not anticipate. Without it, the model picks at random, and you will never know it made a choice.
Role & Audience
Who is speaking, and to whom
The role changes the model’s behavior — it is not decoration. And the audience belongs in the same block: the same content addressed to leadership, a shop floor, or a candidate needs neither the same detail nor the same vocabulary.
Data
With which facts
This is where quality is decided. The other four blocks are framing; this one is the material. It is also the block that clarifying questions fill in, one at a time.
Constraints
What is not negotiable
The legal framework, the scope, the tone, security. And, at the top of the list, the ban on inventing anything: what is missing gets written as missing, never replaced with a plausible value.
Format
The exact expected shape
Sections, length, required mentions, destination medium. A deliverable that arrives in the wrong shape gets redone entirely: format is not the last step, it is a constraint set at the start.
Beyond five blocks
Some deliverables call for more: a few extra blocks get added to the five, never replacing them. They make the tool reason before concluding, and have it review its own work.
The deciding factor is not the expected length of the text, it is the consequence of an error. A three-line letter with legal weight deserves them. A ten-page purely descriptive report does not need them. That is also what keeps everyone from carrying extra weight just to reassure a few.
Four Swiss safeguards, applied by default
A frame that does not say what is forbidden is not a frame. On any sensitive deliverable, these four apply without anyone having to ask.
Never invent anything
Never a legal reference, a labor agreement, a diploma, or a figure that was not given. Missing information gets written as missing, with a visible marker that review cannot miss.
Google notes that forcing structure onto the model "forces the model to create a structure and limit hallucinations." The Template does exactly that, but on the request side.
nLPD · minimization
No personal data beyond what is strictly necessary, and nothing identifying on any medium meant to leave the company.
Cyberhaven’s 2026 report measures that 39.7% of interactions with an AI tool contain sensitive data, and that the average employee does this once every three days. That is not an incident: it is the normal state of affairs.
LEg · non-discrimination
On any document touching employment: no criteria based on age, sex, origin, marital status, religion, health, or orientation. Gender-neutral phrasing.
The check does not happen at review — it happens in the constraints, before the first line is written. A posting corrected after the fact almost always keeps traces of its first version.
Sovereignty
Sensitive data stays on an approved tool. Everything else is Shadow AI, regardless of the good intentions of whoever is using it.
Cyberhaven measures that one in three employees access AI from a personal account, and that 82% of the hundred most-used AI apps are rated medium, high, or critical risk.
Swiss labor law and the Code of Obligations are cited when relevant, and never approximated. A legal basis we are not certain of is flagged for verification rather than guessed at: compliance is not a cost, it is a safeguard.
Sources: Cyberhaven, 2026 AI Adoption & Risk Report, p. 4 and 15 · Lee Boonstra, Prompt Engineering, Google, September 2024, p. 21.
Learn, run, copy
The Method is not bought as a single block. It has three entry points, depending on what you need today, and all three rest on the same five-block Template.
Learn
Understand it, and pass it on internally
"The Augmented Employee" training devotes a full module to the Method and the Template, then twelve hands-on exercises, Template in hand.
Run
Produce a deliverable right now
A tool that starts by asking the missing questions, one at a time, then hands back the complete prompt and the deliverable. It is public.
Copy
Move fast on an already-framed case
The library of role-based prompts, already written to the Template, safeguards included. That is Chaptal.
The Method is published, so it can be checked
The tool that runs the Method is public on GitHub: PALAMBUR/palambur-method. You will find the clarification discipline, the Template, the Swiss safeguards, and adaptations for other work environments.
It does not draft right away: it starts by asking you the missing questions, one at a time, then hands back two things — the complete prompt, ready to reuse, and the deliverable that comes from it. You keep the first one; that is the one with lasting value, because it can redo the second as many times as needed.
A method you cannot read cannot be checked
You can inspect exactly what the tool does with your requests, line by line, before trusting it with anything.
The value is not the file
The framework reads in under an hour. Applying it to your work, your documents, and Swiss law is the actual work — and that is what we sell.
Being public forces us to justify it
A rule written for others is a rule you have to defend. The ones that do not hold up in front of an outside reader do not survive long.
Chaptal: 720 role-based prompts, every one written to the Template
A framework is judged by what it produces. Chaptal is the map of the Method’s role-based prompts: you drill down from your industry to the exact task, and walk away with a prompt ready to copy.
- 720 prompts, 24 roles · from the executive suite to the shop floor, from HR to local government. As of August 20, 2026, and the corpus keeps growing.
- The same five blocks in every one · none of them is a shortcut. Swiss safeguards included, with a visible marker wherever it is your turn to fill something in.
- 24 prompts are open with no account, one per role: enough to judge the quality before deciding anything.
Foire aux questions
How is the Palambur Method different from a simple list of good prompts?
A list of prompts ages along with the models and only covers the cases it anticipated. The Method is a discipline: clarify what is missing before producing anything, frame the request in five blocks, and forbid invention. It applies to any deliverable, including ones nobody wrote a prompt for in advance.
Do you need a technical background to apply the Template?
No. The five blocks are business questions, not IT ones: what needs to get produced and why, who is speaking and to whom, with which facts, under what constraints, in what shape. An HR manager or a shop supervisor fills them in better than an IT person, because they know the answers.
Does the Method work with any AI tool?
Yes — it governs the request, not the tool. It was written and tested on mainstream models, and it holds just as well on a sovereign tool hosted in Switzerland — that is actually the recommended setup as soon as sensitive data enters the request.
Does AI replace human review once the Method is applied?
No, and we will never promise that. The framework sharply reduces guesswork and invention; it does not eliminate them. On any document that commits the company, human sign-off stays mandatory. That is a rule of the Method, not a case-by-case precaution.
Why publish the Method if it is your know-how?
Because a method you cannot read cannot be verified. The framework reads in an hour; applying it to your work, your documents, and Swiss law is the real work. Being public is also what forces us to justify every rule we write into it.
What now?
Reading the Template takes ten minutes. Installing it in a team’s habits takes half a day, and that is exactly what the training does. If you would rather start by measuring what your team already uses, the audit comes first.