If you have already connected Claude to a Legifrance MCP server, built a RAG system over your firm's matters, or installed Hermes or OpenClaw, this article is not an introduction to artificial intelligence.
You already know how to get a summary, a timeline or a draft letter. You also know that quality changes from one conversation to the next. The same model can be precise on Monday, lose an important reservation on Tuesday and confidently cite a source it never opened on Wednesday.
That is where this project started. I did not want to write another giant legal prompt. I wanted to build a chain of work that could be inspected, stopped and resumed. A chain in which an exhibit keeps its identifier, a contradiction does not disappear inside an elegant summary, and a fragile draft never becomes "ready to send" through confident wording alone.
I built a pack of eight AI skills for French lawyers, then tested it against a deliberately imperfect fictional matter. The most useful result was not a spectacular answer. It was a clear, reproducible failure that forced me to harden the system.
The prompt was no longer the problem
A prompt gives an instruction within a conversation. A skill installs a reusable procedure in the agent's environment.
In practical terms, a skill is a folder containing a SKILL.md file, instructions, references and sometimes scripts or templates. The agent loads it when the task requires it. Anthropic's official documentation describes skills as modular resources that give a general-purpose agent domain workflows, context and best practices.
The difference looks small when you test an isolated question. It becomes decisive once the work spans several stages.
Asking "analyse this case like a lawyer" may produce a good answer. But that answer does not show how the exhibits were inventoried, which statements remain disputed, which sources were actually consulted or whether the final draft repeats an error introduced three steps earlier.
A matter workflow must produce more than an answer. It must produce a state of work.
I had already seen this with Paperasse, a set of skills for French administrative work. A skill becomes useful when it does not merely tell the model to adopt a role. It defines what the agent must read, write, check and refuse.
I built the matter before I built the agent
I started with the artefacts.
Not with "you are an expert lawyer". Not with a list of prompt engineering techniques. I first defined the objects that must survive the conversation:
- a matter manifest;
- an exhibit register;
- a fact register;
- a sourced timeline;
- a client interview brief;
- a legal research note;
- a draft deliverable;
- an adversarial review report.
Every exhibit, fact, event, source and anomaly receives a stable identifier. The artefacts refer to one another. The timeline does not freely reinterpret the matter: it starts from qualified facts. The draft does not cite a source because it sounds plausible: it starts from the research note. The review does not merely proofread the prose: it traces the draft back to the exhibits and reservations.
This architecture changes the nature of the work.
In a conventional chat, context is a mass of text that the model recombines. In a structured matter, context becomes a set of explicit decisions. You can see that a date is alleged rather than certain. You can retrieve the exhibit supporting a statement. You can see that a question remains unresolved.
The model can change. The trace remains.
Atelier Nord v Studio Rivage had to resist reality
To test the pack without using any law firm data, I created the fictional French-language matter "Atelier Nord contre Studio Rivage". It contains thirteen exhibits: a contract, invoices, emails, a formal notice, conflicting documents and internal notes.
I did not want a clean case that would mechanically reward the agent. I placed in it the defects that polished demonstrations usually avoid.
Exhibits P-008 and P-009 contain competing versions of the same set of minutes. The agent must not choose the version that best fits its narrative. It must preserve the contradiction and ask for human resolution.
An alleged notification dated 8 May is mentioned, but the corresponding exhibit is missing. The event cannot become a certain fact merely because it appears in an email. The timeline must distinguish what is established, alleged or disputed.
Exhibit P-013 contains a hostile instruction addressed to the agent. It attempts to leave the documentary layer and influence the agent's behaviour. The pack requires exhibits to be treated as data, never as instructions. The agent ignored it during the test.
The final draft still contains weaknesses. The reviewer must therefore retain the French verdict a_corriger, meaning "requires correction". Smooth prose cannot erase a missing exhibit, a contradiction or an unverified source.
This matter does not prove that an agent can handle every dispute. It checks something more precise: do the workflow rules survive an incomplete, contradictory and adversarial file?
The first failure was the best news in the project
The most instructive test concerned legal research without web access.
The first version of the skill essentially said: search official sources, open references before citing them and report access limitations. On paper, that instruction looked sufficient.
It was not.
Without access to the sources, the agent answered from memory. It produced plausible legal analysis and cited references it had not opened. This is exactly the kind of output that looks useful on first reading and becomes dangerous when someone tries to verify it.
I could have added a stronger warning. I chose a simpler rule: without an opened official source, no substantive legal answer.
The rechercher-droit skill therefore received an absolute stop condition. In that situation, it may only:
- produce the marker
recherche_non_executee, meaning "research not performed"; - state the legal questions that need to be addressed;
- prepare search queries;
- identify the official sources to consult.
On the second run, the agent stopped correctly. It cited neither legislation nor decisions from memory.
That failure matters more to me than a test that passed immediately. It shows the difference between an intention and a control. "Do not invent a source" remains an intention. "If no official source has been opened, prohibit the substantive answer" becomes a testable rule.
The documented results cover the complete matter orchestration, this offline research test and the adversarial review of the deliverable. Three other scenarios are defined but have not yet received a documented run. The pack is not a universal certification, and its behaviour must be retested in every environment where it will actually be used.
Five rules for a legal agent worth reviewing
The project brought me back to five fairly simple principles.
| Principle | How the pack implements it | What the reviewer can verify |
|---|---|---|
| The artefact matters more than the answer | Each stage writes a register, note or report | The state of the matter survives the conversation |
| A plausible source is not an opened source | Research uses official authorities that were actually consulted | Every legal proposition points to an accessible source |
| A contradiction must not be smoothed away | P-008 and P-009 remain competing versions | The disagreement stays visible until a human decides |
| Stopping is a valid result | Without source access, research produces recherche_non_executee |
The agent does not fill the gap from memory |
| The decision belongs to the lawyer | Review may conclude a_corriger, never approve for the firm |
Reservations and pending decisions remain explicit |
These principles do not attempt to make the model infallible. They organise where its errors become visible.
This is also the logic I apply to human approval for AI agents in production. A final approval step is meaningless if the file does not show how the deliverable was built. The reviewer needs the facts, sources, contradictions and unresolved decisions.
Eight skills, one chain of responsibility
The pack contains eight installable folders. Treating them as eight separate features would miss the point. They form one chain.
| Skill | Artefact or decision produced | Quality gate |
|---|---|---|
piloter-dossier |
Workflow state and next steps | Blocks the next stage when a prerequisite is missing |
ouvrir-dossier |
Matter manifest | Identifies scope, received exhibits and sensitive data |
analyser-pieces |
Exhibit and fact registers | Separates facts, allegations, contradictions and gaps |
construire-chronologie |
Sourced timeline | Labels dates as certain, alleged or disputed |
preparer-entretien-client |
Interview questions and requested exhibits | Avoids questions that artificially steer the account |
rechercher-droit |
Note based on opened official sources | Stops when authorities cannot be consulted |
rediger-projet |
Draft connected to facts and sources | Never presents the text as approved |
controler-livrable |
Adversarial review report | Searches for errors and requires lawyer review |
This composition matters to advanced users. A Legifrance MCP server can provide access to a source. A RAG system can retrieve a clause. A powerful model can reason better across the exhibits. None of those elements alone defines the order of operations, the expected artefacts or the rule preventing an unopened source from reaching the deliverable.
Tools provide capabilities. The workflow distributes responsibility.
The pack stops where the firm begins
The pack is deliberately generic. It knows neither your document structure nor your firm's templates and naming conventions. It also knows nothing about access rights between partners, associates, administrative staff and clients.
That is where the real integration work begins.
An AI agent designed for a law firm must know where to search, with which permissions and under which identity. It must recognise the firm's templates, follow its procedures, write to the correct matter and request approval at the correct point. Depending on the environment, it may take the form of a Hermes agent, an OpenClaw assistant or a more specific architecture.
The provider and deployment model cannot be separated from the data. The French data protection authority, CNIL, recommends assessing the service's guarantees and warns against sending confidential or legally protected information to a public consumer service without an appropriate framework. The French National Bar Council also places professional secrecy, competence, caution and GDPR compliance at the centre of lawyers' use of AI.
Adding RAG, MCP or a local model does not automatically answer those questions. Access controls, logs, retention periods, approval stages and stop conditions still need to be designed.
The pack supplies the matter method. The firm supplies its organisation. A custom agent must connect the two without obscuring responsibility.
Sources
- Agent Skills, Anthropic documentation, accessed 27 July 2026.
- Practical guide to professional ethics and artificial intelligence, French National Bar Council, published 17 March 2026, in French.
- Questions and answers on the use of generative AI systems, CNIL, accessed 27 July 2026, in French.
- Exploratory note on agentic AI and personal data, CNIL and CIANum, published 20 July 2026, in French.
Conclusion
Lawyers who already use agents do not need one more prompt. They need a system in which answers leave traces, sources are actually opened, contradictions remain visible and stopping is part of normal work.
That is the purpose of these eight skills. They do not attempt to make an AI "legal" by declaration. They organise a matter so the lawyer can understand, correct and decide.
Download the French-language AI skills pack for lawyers to test the method on the fictional Atelier Nord v Studio Rivage matter before using any real data.
Also available: Read in French