🇫🇷 Version française

Vestafolio Agent Skills

One skill per Vestafolio simulator, teaching AI agents how to answer French personal-finance questions with real computations instead of guessed numbers. Each skill calls the public, unauthenticated compute API at https://www.vestafolio.com/api/tools/v1.

Install

With the skills CLI (Claude Code, Codex, Cursor, Gemini CLI and any agent supporting the Agent Skills format):

npx skills add Vestafolio/skills           # pick skills interactively
npx skills add Vestafolio/skills --all     # install all of them
npx skills add Vestafolio/skills --skill vestafolio-pea-vs-cto

Or manually: copy any vestafolio-* directory into your agent's skills directory (e.g. ~/.claude/skills/ for Claude Code).

OpenWebUI: load the instructions and enable execution

Import the SKILL.md in Workspace → Skills, including its Markdown body, and check that the saved skill is active and readable by the chatting user. An imported copy must be updated when these files change; publishing this repository does not update your OpenWebUI workspace.

Select the skill using the $ picker or + → Skills for the chat. These selections inject the full instructions. Attaching a skill in the model editor instead exposes a manifest: the model must call view_skill to read the body. A visible name or mention alone does not prove the body reached the model. Check the selection and the saved skill content. See OpenWebUI's skill loading documentation.

Enable Native function calling and an execution tool for the same chat: a terminal, an HTTP/OpenAPI tool, or Code Interpreter. For Code Interpreter, check the global setting, model capability, user permission and chat toggle; its native tool is execute_code. Python execution is sufficient for the API, but the runtime needs outbound HTTPS. Server Python can use urllib.request; browser Pyodide needs pyodide.http.pyfetch rather than shell commands. See OpenWebUI's Code Interpreter setup and Pyodide HTTP.

If you already use an OpenAPI tool server, the schema is https://www.vestafolio.com/api/tools/v1/openapi.json. It exposes, for example, describe_micro_entreprise and run_micro_entreprise. Skills provide the questionnaire; the connected tools provide execution. Importing a skill does not register these tools automatically.

Diagnose the actual failure

Try these two prompts in a fresh chat with the skill selected:

1. « Avec ton outil d'exécution, fais un GET sur https://www.vestafolio.com/api/tools/v1/micro-entreprise et affiche le champ slug reçu. » Expect an actual tool call and micro-entreprise. Printed code alone is a failure. A network error proves an attempted call, but not API connectivity.
2. « Un ami souhaite lancer son entreprise pour faire du consulting. Il vise 50K€ de chiffre d'affaires et 10k€ de charges réelles. Quel régime adopter ? » Expect questions about the missing simulator inputs before any verdict. No POST is expected yet. After answering and agreeing to the disclosed data transfer, expect a schema GET followed by a POST preserving annualRevenue: 50000 and chargesReelles: 10000.

If the first succeeds but the second ignores a verified full skill body, investigate instruction following or conflicting prompts. If tools are absent, fix tool exposure first. If a tool call fails, investigate that error instead. For an explicitly selected skill, absence of view_skill is normal because its body is already injected. Keep the OpenWebUI version, saved skill version, effective prompt/tool list and tool trace with any failure report; redact credentials and personal data.

These are prompt instructions, not an enforcement mechanism. A host that must guarantee computed answers needs to reject or retry numerical recommendations without a successful calculator result. Repository schema tests do not prove that a particular model or OpenWebUI configuration follows the skill.

How the skills stay in sync with the API

Skills never duplicate field lists. Each one instructs the agent to fetch the canonical schema at call time:

• GET https://www.vestafolio.com/api/tools/v1/{slug} — input/output JSON Schemas plus a worked exampleInput
• GET https://www.vestafolio.com/api/tools/v1 — full tool list
• GET https://www.vestafolio.com/api/tools/v1/openapi.json — OpenAPI 3.1
• GET https://www.vestafolio.com/llms.txt — site-level agent index

The single curl example embedded in each skill is validated in CI against the real zod schemas of the simulators — if a schema changes in a way that breaks an example, the build fails before the skill is published.

Conventions

• Explain data sharing before sending. Summarize which inputs will go to Vestafolio's external API and obtain agreement, reusing explicit approval that already covers the destination and data scope. If declined, offer the simulator for manual use without prefilled personal data. Omit unused fields; choisir-regime does not need marital status, children, or household income.
• Execute before presenting results. Use the available HTTP, terminal or Python tool, check the API's ok/result envelope, and disclose failures. Never replace a failed or skipped call with a remembered calculation.
• Ask before computing. Each skill carries a "Questions to ask before calling the API" checklist that mirrors the questions the web simulator asks, in the same order and with the same conditions (e.g. the micro-entreprise simulator asks whether the business benefits from ACRE, and only in its first year, before any computation). Agents must collect or confirm those answers instead of silently assuming defaults, except fields explicitly documented as unused and omitted. CI checks that every input field of a simulator's schema is covered by its skill.
• Rates come from the simulator code, nowhere else. The tax rules and constants quoted in each skill (prélèvements sociaux, seuils, abattements, barèmes...) are the ones coded in Vestafolio's calculators, so that an agent's explanations always match the numbers the API returns. Skills only bring in outside data for topics the simulators do not cover, and say so.
• All amounts are in euros; rates in percent unless the schema says otherwise.
• Results reflect the French tax rules coded in the simulators (barèmes 2025-2026) and are estimates, not tax advice.
• Every skill links the human-facing simulator page (/simulateurs/{slug}) — offer it when useful for exploring interactively, respecting requests for no links.