English · Français · Deutsch · Italiano
A skill that asks the right questions before it writes. You ask an AI for a working document (a job advert, a quote, a payment reminder, a contract, an internal memo). Instead of guessing what is missing and making up the rest, the skill asks you at most four short questions, one at a time, then delivers a document you can actually send. Designed for businesses in French-speaking Switzerland, usable anywhere. Free, MIT licence.
The skill itself is written in French · it produces French-language documents for businesses in French-speaking Switzerland, and its published measurements were made on that text. What a visitor reads before installing (this README, the contribution guide, the installation guide) is in English.
Claude Code · two commands in the conversation, then one request.
/plugin marketplace add PALAMBUR/palambur-method
/plugin install palambur@palambur
Then type: palambur method: write a job advert for a painter at 80 %
In your browser (Claude.ai, ChatGPT, Infomaniak’s Euria, Mistral…) · open adaptations/methode-palambur-prompt.md, copy the block below the rule, paste it into the instructions of your project or assistant. Nothing to install. For an instructions box with a tight size limit, take the short prompt. Both prompts are in French; the questions and the deliverable follow the language of your request.
Codex, Cursor, Copilot, Gemini CLI, Mistral Vibe · a single symbolic link from the skills/palambur-method/ folder into ~/.agents/skills/ serves all five hosts. The commands, host by host, are in docs/portabilite/INSTALLATION.md.
Check that the skill is loaded · in Claude Code, /plugin then the Installed tab; or, on any host, ask “palambur method: what is your first question?” and expect a reply that starts with “Question 1 of N”.
- Try it in two minutes
- Before and after, in 30 seconds
- What it is
- Triggering · three doors
- Installation
- Repository layout
- Measures and test cases
- Commitments
- Contributing · Citing · Licence
What you type:
palambur method: write a job advert for a painter at 80 %
What the skill replies. On its first trigger, three opening lines, the sensitivity level, then the first question, all in the same message:
Palambur Method. I will first ask you a few questions, one at a time, so that I invent nothing on your behalf. You answer with a letter or in free text; if you skip, a cautious default applies and stays visible. You will then receive two blocks to copy: the reusable Palambur Template, then the Deliverable.
Sensitivity level · ordinary (no name, no amount)
Question 1 of 4 · the place and the setting
Where is the position, and in what setting?
A. a single workplace, canton to be named B. several sites or regular travel C. partial remote work possible, share to be specified D. Other / I don’t know · default if you don’t say
Then, one question per message, always with a counter, lettered choices, a flagged cautious default and an escape hatch:
Question 2 of 4 · the duties · A. I will provide them · B. propose the usual duties of the trade, I will correct · default if you don’t say · C. the position is new, the duties are not settled yet · D. Other / I don’t know
Question 3 of 4 · the profile · A. a required qualification (CFC, federal diploma, master craftsman) · B. CFC or recognised equivalent experience · default if you don’t say · C. no qualification required, in-house training provided · D. Other / I don’t know
Question 4 of 4 · what the company offers · A. nothing beyond the contractual and collective framework · default if you don’t say · B. benefits to be named · C. benefits to be named and a published gross salary range · D. Other / I don’t know
Answers kept for the example: A (canton of Vaud, workshop in Yverdon-les-Bains), B, A (Swiss federal certificate, CFC, in building painting), B (13th month’s salary and continuing education). If you choose “I don’t know”, the field stays [TO BE COMPLETED] in the document.
What you receive. First one line of context, stating what was kept:
Context kept · job advert · building painter at 80 % · canton of Vaud · factual tone · data: none · 4 answers, 0 defaults, 10 markers
Then the completed Palambur Template, the reusable prompt (excerpt):
Palambur Template
INSTRUCTION
Write a complete job advert for a building painter position at 80 %, to be
published on [TO BE COMPLETED: channels]. Goal: attract qualified applicants
by stating the real work, the place, the hours and the setting, without
promising anything the management has not validated.
ROLE AND AUDIENCE
You are the recruiter at Exemple SA, a building painting company in the
canton of Vaud, trained in the Code of Obligations, the Gender Equality Act,
the Labour Act and the Data Protection Act. You write for a tradesperson who
reads the advert on a phone.
DATA
• Job title: building painter (m/f/d)
• Workplace and canton: workshop in Yverdon-les-Bains, sites across the
canton of Vaud
• Rate · contract type · start date: 80 % · open-ended · to be agreed
• Required profile: CFC in building painting
• Benefits validated by management: 13th month · continuing education
• Applicable collective labour agreement, if known: [TO BE COMPLETED]
CONSTRAINTS
No discriminatory criterion (Gender Equality Act, Code of Obligations),
gender-neutral wording (m/f/d). No invented salary, no invented holiday
entitlement. No legal article quoted: if a collective agreement applies,
name it and write [TO BE VERIFIED: applicable agreement].
FORMAT
Six sections: title and hook · the company · your duties · your profile ·
what we offer · how to apply. 400 to 600 words. Formal address, French of
French-speaking Switzerland. Data processing notice at the end of the advert.
Then the Deliverable (excerpt):
Building painter, CFC (m/f/d) · 80 % · Yverdon-les-Bains
Do you take pride in careful work and renovation projects? Exemple SA is looking for a building painter at 80 %, start date to be agreed.
Your duties [TO BE VALIDATED: proposed list, to be corrected by management] · surface preparation, interior painting of flats and offices, wall coverings, finishing, on-site advice to clients.
Your profile · required: CFC in building painting · desirable: [TO BE COMPLETED: years of experience], driving licence [TO BE COMPLETED: category].
What we offer · 13th month’s salary · continuing education paid for.
How to apply · send your file to [TO BE COMPLETED: recipient’s function] via [TO BE COMPLETED: channel].
Your application data is processed by Exemple SA using [TO BE COMPLETED: tool], hosted in [TO BE COMPLETED: country], and kept for [TO BE COMPLETED: duration].
And, to close, the Trust log, five lines stating what was protected, marked and refused:
Trust log
· Sensitivity level · ordinary (no person named)
· Data · none
· Markers placed · [TO BE COMPLETED] ×7 · [TO BE VERIFIED] ×2 (applicable agreement, job posting notice) · [TO BE VALIDATED] ×1 (duties)
· Refused assumptions · no salary figure; no collective agreement named from memory; no data hosting location asserted
· Human review · required before publication (HR deliverable)
Whatever was not said has not been invented: it stays marked. If your request contains client names or amounts, the skill first asks, in one question on its own, whether it should anonymise them (anonymising is the default). The full exchange is in docs/exemples/01-offre-emploi.md; nine other walkthroughs, one per template, are in docs/exemples/ (in French).
Two names, two objects. The Palambur Method is the whole: the discipline of clarifying first, the Swiss legal guardrails, the refusal to invent. The Palambur Template is its tool: the 5 ordered blocks of a prompt, Instruction · Role and audience · Data · Constraints · Format, plus 4 extension blocks for high-stakes deliverables (contracts, charters, internal policies).
The skill turns an under-specified business request into a quality deliverable by filling the missing context before producing anything. Internally it anticipates a batch of at most four decisive questions (place and setting, duties, profile, benefits…), but asks only one per message, with a counter (“Question 1 of 4”), a flagged cautious default and an escape hatch (“Other / I don’t know”). Then it delivers:
- the completed Palambur Template, ready to copy, paste and reuse;
- the written Deliverable, with the Swiss guardrails (Gender Equality Act, Data Protection Act, Code of Obligations, Labour Act, data sovereignty), closed by a Trust log.
Three commitments.
- Smooth · few questions, well targeted, four per batch at most. If everything is already provided, it produces straight away.
- Safe · data protection by default (anonymisation, data protection notice), no invented data, no legal article number written without a
[TO BE VERIFIED]marker. - Silent · this skill sends nothing, anywhere. No telemetry, no Palambur server, no account. It is plain text files that your AI reads.
A company profile, so you never answer twice. The skill reads a profil-entreprise.md file placed at the root of your working folder: canton, collective agreement, tone, signature, language. Its completed fields count as answers, and the questions it settles are no longer asked. The model to copy is in the skill: skills/palambur-method/profil-entreprise.md.
And a review mode. You attach a text you have already written (a dismissal letter, a quote) and ask for a review: no questions, a table “Finding · Risk · Proposed correction”, then the corrected version. The review fixes form, law and data; it does not rewrite the substance (relecture.md).
One is enough.
- The word · your message contains “palambur”, “méthode palambur” or “palambur method”.
- The nature of the request · a working document asked for in one or two sentences (job advert, quote, reminder, contract, internal memo…), a Swiss business context, names or amounts to protect, or a text attached for review.
- The “always on” option · write “Méthode Palambur toujours active” in the conversation or in your
CLAUDE.md, and the skill applies to everything the AI writes, until “Méthode Palambur désactivée”.
Once triggered: if decisive elements are missing, it clarifies; if everything is provided, it produces directly; if a text is attached for review, it reviews without asking anything.
The skill’s own files are written in French. The questions and the deliverable follow the language of your request.
- To use the skill · a host that reads the
SKILL.mdformat (an up-to-date Claude Code, Codex, Cursor, Copilot, Gemini CLI, Mistral Vibe) or any AI with an instructions box. Nothing else. - For the repository tools (control chain, test cases, generators) · Python 3.9 or newer. Optional: you do not need it to use the skill.
In Claude Code, add this repository as a marketplace:
/plugin marketplace add PALAMBUR/palambur-method
Then install the plugin:
/plugin install palambur@palambur
Updates arrive with each new version of the plugin; /plugin then Installed shows the one you have.
If you prefer a copy in ~/.claude/skills/ without the marketplace, clone the repository and run the script. Dry run by default, it shows what it would do:
bash outils/installer_local.shWith --ecrire, it actually writes the copy:
bash outils/installer_local.sh --ecrireThe script refuses if the target is already a symbolic link, so as not to write into another clone through the link.
The SKILL.md format is an open standard. A single symbolic link from skills/palambur-method/ into ~/.agents/skills/ serves five hosts: Codex (which follows symbolic links and also reads .agents/skills/ inside a project), Cursor, GitHub Copilot, Gemini CLI and Mistral Vibe CLI (each vendor’s official documentation, read on 11.10.2026). The Gemini app and Mistral Vibe Work accept the skill by upload. The step-by-step guide, host by host, with its dated sources, is in docs/portabilite/INSTALLATION.md; the skill’s conformity to the standard is recorded in docs/portabilite/CONFORMITE_STANDARD.md (both in French).
For hosts that only offer a system instructions box, the adaptations/ folder provides the method condensed into a standalone prompt, generated from the skill itself at each version: a full version under 8'000 characters and a short version under 1'500 characters. On these hosts the assistant is the method: it applies to every request, with no trigger word. The table “which host, which version, where to paste” is in adaptations/README.md.
One truth, the repository. The skill has a single source: skills/palambur-method/ in this repository. What the hosts read (~/.claude/skills/, ~/.agents/skills/, an installed plugin, a pasted prompt) is a link or a dated copy, updated with git pull or a new plugin version, never by editing the files in place: a fix made in a copy is lost at the next update, and nobody reviews it.
palambur-method/
├── README.md · README.fr.md · README.de.md · README.it.md # this file, in four languages
├── CHANGELOG.md · CONTRIBUTING.md · CODE_OF_CONDUCT.md # versions, contributions, conduct
├── SECURITY.md · CITATION.cff · LICENSE # reporting, citation, MIT
├── .claude-plugin/ # plugin.json and marketplace.json · the Claude Code plugin
├── .github/ # issue templates, PR template, workflows/controle.yml (CI)
├── .agents/skills/palambur-method/ # Codex and .agents hosts location · a README, the link is made on your side
├── skills/palambur-method/ # THE SKILL · the single source
│ ├── SKILL.md # when to trigger, steps 0 to 4, the shape of the questions
│ ├── methode-palambur.md # the Template (5 + 4 blocks), the 4 levels, the tone, three walkthroughs
│ ├── protection-donnees.md # sensitivity levels, anonymisation, trust log
│ ├── relecture.md # the review mode · six screens, output table
│ ├── references-juridiques.md # the legal texts the skill may name, what it never invents
│ ├── profil-entreprise.md # the company profile model read at the root of the working folder
│ ├── gabarits/ # one file per frequent deliverable (questions, DATA fields, pitfalls)
│ ├── metiers/ # one file per trade (vocabulary, named legal texts, pitfalls)
│ └── agents/openai.yaml # optional metadata read by Codex
├── adaptations/ # the generated system prompts · full (under 8'000 characters) and short
├── docs/
│ ├── exemples/ # ten complete walkthroughs, one per template, fictional data
│ ├── pedagogie/ # MANIFESTE.md, FAQ.md, GLOSSAIRE.md
│ ├── portabilite/ # INSTALLATION.md (host by host), CONFORMITE_STANDARD.md
│ ├── VEILLE_JURIDIQUE.md # what feeds the table of legal references
│ └── DIX_COMMANDEMENTS.md # data protection in ten lines
├── epreuves/ # 149 cases (cas.json), grille.md, mesurer.py, generer_evals.py,
│ # jouer_apertus.py, results and protocol, reponses/ (transcripts)
├── evals/ # README only · the “claude plugin eval” cases are regenerated
└── outils/ # controle.py, generer_adaptations.py, installer_local.sh, regle_typo.py, metiers/
The gabarits/ and metiers/ folders grow with each version; the changelog says what is in place, with a quality note per version. The ten commandments hold data protection on one page; the FAQ, the manifesto and the glossary answer the recurring questions. The documentation outside this README is in French; the German and Italian READMEs await a native review.
The control chain checks the repository before every publication: typography rule, forbidden word, valid JSON, a single version number across three files, adaptations up to date with their sources, SKILL.md frontmatter and size, relative links. It runs on any clone and in the CI (.github/workflows/controle.yml), with no secret and no model call:
python3 outils/controle.pyThe control breaks itself on fabricated witnesses, to prove it is looking:
python3 outils/controle.py --epreuveThe skill’s test cases (is it smooth, is it safe?) are in epreuves/: 149 cases in three families (deliverables, trade tasks, traps that push the model to invent · as of 11.10.2026, python3 epreuves/mesurer.py --compter is authoritative), a marking grid with twenty guards, and mesurer.py, which marks a reply against the grid and can break itself too (--epreuve). The same cases in claude plugin eval format are regenerated from epreuves/cas.json with python3 epreuves/generer_evals.py; they are no longer versioned, and evals/README.md says how to run them.
The published measure. Measured on 12.10.2026 on Apertus (Infomaniak) with the full adaptation prompt, not the native skill: 1 green out of 149. The model asks the anonymisation question, then delivers the Template and the deliverable in the same message, without clarifying and without markers. Details, red guards per family and exact spend are in epreuves/RESULTATS_APERTUS_2026-10-11.md (in French). The native skill on Claude, measured on 13.10.2026 with claude plugin eval (Claude Code 2.1.296, model claude-opus-5-5), first turn only, one run per case, on the ten cases of the three-model protocol: 8 green out of 10 with the skill, 0 out of 10 without it (same model, same prompts, plugin removed). The two reds are the instrument’s, not the Method’s: on reading, both first turns follow the Method, one names the fictitious law inside its refusal, the other proposes « Client X » without the word « anonymise ». Without the skill the model writes the whole document at once, cites a law article and an interest rate bare, and drifts to French law in three Swiss cases. Transcripts of both arms, estimated spend and limits of the measure are in epreuves/RESULTATS_CLAUDE_NATIF_2026-10-13.md (in French).
- The base skill is free and will stay free. MIT licence: you may use, copy and modify it, including commercially. The base skill is this whole repository: the method, the templates, the trade files, the test cases, and every version to come. Palambur sells service around it (a version pre-configured for one company, an assisted installation); nothing that is here becomes paid.
- It sends nothing, anywhere. No telemetry, no network calls, no account. Your requests stay between you and the AI you chose.
- No advertising inside the skill. Deliverables carry no Palambur mention. The skill names one single Palambur tool, Jomini (free anonymisation in the browser, no account), because it serves data protection; that is the only mention. What follows is the other exception, and it lives here, not in the skill.
The skill is the door, Chaptal is the house: 782 ready-to-use prompts per trade, built on the Palambur Template, on palambur.ch (as of 11.10.2026).
A missing template, a legal correction, a new language: CONTRIBUTING.md says how to propose one, what the control chain requires and the native speaker rule for translations. The code of conduct applies to every exchange; a vulnerability, or a behaviour of the skill that exposes data, is reported as described in SECURITY.md, not in a public issue.
To cite the Palambur Method in a paper, a course or an article, the citation metadata is in CITATION.cff; GitHub offers its APA and BibTeX forms from the repository page.
MIT · Palambur · Victor Semblanet · 2026