1,081 words, about 5 minutes.
In 1010, Emperor Lý Thái Tổ moved his capital to a bend of the Red River where, the chronicle says, a golden dragon rose from the water into the sky. He named the city Thăng Long, the ascending dragon. I spent years of my research career in that city, now called Hanoi, and the dragon I met there is not the one I grew up with in France.
The European dragon is a monster. It hoards gold, burns villages, and exists so that a knight can kill it. The East Asian dragon is something else: a power of nature. It brings the rain that feeds the rice fields, it commands the rivers, and it stands for the emperor himself. Nobody slays it. The question it poses is not how to kill it but how to live with something that much stronger than you.
That is how I think about AI assistants in research.
An AI assistant has enormous capability and no judgment. It will draft a literature review in a minute and invent three of the references. It will summarize a paper it has not read. It will refactor your data pipeline and silently drop the step you forgot to mention. None of this is malice; a river has no malice either. The damage comes from treating a power of nature as if it were a colleague who checks their own work.
Researchers can too easily fall into extreme attitudes. The knights want to slay the dragon: ban the tools from the lab, the journal, the classroom. They will lose, because the productivity gap is already too large to ignore. The worshippers copy whatever the oracle produces into their manuscripts. They will lose too, and more publicly: a fabricated citation in a published paper stays fabricated forever. There is a third way, the high road: learn to ride the dragon, make your harness.
A harness is a set of equipment that turns a beast more powerful than you into a productive partner that works in your direction. For the doctoral students, I wrote a book of advice, “Research excellence à la Française: How-Tos for Students and Professionals” (REALF, get your free copy on my homepage!). Claude and ChatGPT need to be given the same material, only in a different way. They need a persistent layer to transform a generalist language model —a system trained to predict the next word, recursively— into an agent adapted to a professional activity.
LLM + Harness = AI Agent
Now, why would a researcher write their own harness, instead of using whatever the AI bros provide? Thanks for asking. For the same reason I use Linux. The big tech corps do everything they can to lock us into their own ecosystems, hoping to monetize that dependence later — part of the dynamic Cory Doctorow calls “enshittification”. Writing my own harness, I know how it works, and I am free to choose whatever LLM provider I want. In particular, a well-designed harness for a specific profession can compensate for less intelligent models (read: work someday with a model I can run on my own workstation). Also, the labs limit what their agents are allowed to do. They side with the editorial cartel in the open-science war, for example.
My harness’s name is the Imperial Dragon Harness (IDH, get your free copy on GitHub!). The name was inspired by Dragon Dreaming, a project-design method developed by John Croft and Vivienne Elanta in Australia, drawing among other sources on Aboriginal Australian knowledge, deep ecology and systems thinking. It posits that projects cycle through four interconnected phases: Dreaming, Planning, Doing and Celebrating. Modern harnesses, like IDH, organize the action loop in five steps: Imagining, Planning, Doing, Verifying and Celebrating. In business-speak: Conception, Strategy, Execution, Evaluation, Closure. In Vietnamese art, the dragon with five sharp claws is reserved for the King, hence the adjective.
So what is a harness, materially? It is a set of instructions including:
1. Rules. They are automatically read by the agent when certain conditions occur. For example, at the start of a conversation the agent reads a file named AGENTS.md or README.md, which contains among other instructions:
Every task passes through five phases:
| Claw | Phase | Activity |
|---|---|---|
| 1 | Imagine | Explore, brainstorm, surface motivations |
| 2 | Plan | Design, write tickets with test specs |
| 3 | Execute | TDD red/green/refactor, open PR |
| 4 | Verify | Review PR, fix, iterate ≤3 cycles |
| 5 | Celebrate | Reflect, consolidate memory, dream forward |
The IDH also contains rules on how to write in academic English (follow the Elements of Style), what to do when releasing a paper, how to find a suitable journal to publish, how to use Zotero, how to write Python the way I like… Many are adapted from the REALF book, since Claude is close to a graduate student.
2. Skills. These are commands that tell the system to do something. In the IDH, the skills are Dragon-themed, using short and memorable verbs. Those I use the most are:
| /hunt N | Look at issue number N. Plan, Write a solution, Verify and fix (no more than 3 rounds). |
| /gaze | Organize a multi-expert review of the solution, using independent AIs, and revise it, leaving no remarks unanswered. |
| /roar | Wrap up after solving an issue: update the documentation, reflect on what worked, verify if similar issues still exist… |
| /raid N, M… | Organize a multi-wave parallel work plan for issues N, M…: Imagine, Plan, Implement, Verify, Deliver and Celebrate them all. |
| /lair | Terminate the day’s work: housekeeping, save unfinished work, run the full test suite, consolidate memories… |
| /perch | Tell me what we have done so far, what remains to be done, and how we can safely stop the conversation here |
There are also more specialized skills for academic work. For example, /biblio-saturation sends independent agents hunting through different disciplines, languages and citation networks to stress-test a claim of novelty; /external-peer-review submits a manuscript to independent AI reviewers and synthesizes their disagreements.
3. Scripts. Because rules and skills are only indicative, and open to interpretation, everything that is algorithmic is better encoded in Python. For example, when an agent finds new references, a script merges them into the project’s BibTeX database, deterministically checking DOIs and titles for duplicates instead of asking the LLM to decide whether two references are the same.
You do not have to build a harness as elaborate as the IDH. But if you work with a dragon, you need one.
PDF version for print: AI is a Dragon
OpenEdition suggests that you cite this post as follows:
Minh Ha-Duong (August 31, 2026). AI is a dragon. Research excellence à la Française. Retrieved October 6, 2026 from https://realf.hypotheses.org/2
Awesome ! For me, I want my own imperial dragon too but I’m not doing research so I guess I’ll have to build my own harness from scratch. My problem is, I did not write a book about how I work so I have nothing to feed my dragon to produce the harness. How many hours you’d say it takes to create my own harness ?
Creating your own harness? In the real world, you bootstrap: use an existing harness to write your specialized harness. In a few hours you can show an agent how to do one task. Then the agent will create the skill for you (and help you debug it). For example I have a 216 lines
AGENTS.mdfile that tells an AI how to send their monthly receipts to my tenants using a linear nine-steps workflow. The stock harnesses from commercial products likeHermes,claude,codexalready knows how to use common tools, manage memory, plan and execute tasks — ready to adapt to your ways. But if you start from a bare bone harness like Pi or ZeroClaw, then it will take a few days to get to the point where it does useful work. I took a seminar two months ago where we wrote a “harness in Python in 2 hours: run a tool-capable model in a loop, call the external tools made available to the model, and build the next prompt using the results. But this runtime is actually more a skeleton than a dragon. The real muscle is in the tools you make available.