Year
2026
Role
Designer and builder, directing the AI and reviewing every result
Context
Independent project
Duration
Oct 2026
Team
Solo, built with Claude Code and n8n
Status
Four prototypes, tested on small or made-up samples
The problem
Research work is full of slow, repeated steps. General AI tools can reword quotes, lose their sources and invent references.
The outcome
Four working prototypes, each with a written test record and a list of what I cannot claim.
4
builds, from a fixed workflow to a team of agents
Three levels of AI autonomy
20
stress tests on one workflow
Insight Synthesis Engine
5
AI specialists in one studio
Two briefs run end to end
Overview
I wanted to know how much to let an AI decide in research work. So I built four things, from a fixed workflow where the AI does one small job, up to a team of agents that starts from a single sentence.
Each one is split the same way: what the AI decides, what fixed rules decide, and what a person checks. I built them in n8n and Claude Code, with me directing and reviewing. They are prototypes, tested on small or made-up samples.
Who decides what, across the four builds.
DISCOVER
Workflow: Insight Synthesis Engine
A researcher pastes interview transcripts, notes and web links into one form. The workflow hides the names they list, sorts the text into themes and writes one insight per theme, with the exact quotes behind it. The quotes are copied by code, so the AI cannot reword them.
The first version needed the user to edit JSON inside a node. My stress test showed that was fragile, and that some failures looked like success. I asked for a rebuild instead of patches: a form in place of JSON, a check after every stage and a plain message when one fails. The test also found that names in Hindi and with accents were not being hidden. They are now.
The pipeline is steady, but the analysis is not. Five runs on the same material placed 14 to 29 of 40 pieces in themes, so one run is a first draft. Accuracy was never compared with human coding.
The canvas, the form, the report and the numbers from a sample run on made-up interviews.
DEVELOP
Agentic AI: Design Team
Five fictional specialists, each with a personality, its own knowledge, its own memory folder and web search, turn one sentence into research, priorities, a design, a review and, if asked, starter code. I required each agent to have a name, memory, experience and knowledge, and held the build to that list. Testing found one agent that saved no memory and another with no index file, so saving memory became a mandatory step.
It first ran on free models in n8n and hit daily limits. I did not want to pay for API access and already had a Claude plan, so the main version moved to Claude Code and the n8n version became the backup.
Two sample briefs ran end to end, in about 10 and 25 minutes, and a hidden instruction planted in one of them was ignored. I have not tested the quality of its design advice with users.
A one-sentence brief passes through five specialists. The characters are made up.
DELIVER
Two smaller builds
Literature Review Screener. An n8n workflow that screens papers for a systematic review, from a topic search or my own list of DOIs, with a reason for each decision and PRISMA counts. I tested it on my own ICWIHI 2025 review. The AI screener kept 14% of my known papers against a target of 80%, and quoted the written criterion each time. The review's written criteria were narrower than what I actually included. I report it as it came out. The Zotero push and a conference branch are not built.
Research Writer. One agent that checks a draft against a quote ledger, verifies references and polishes my own sentences. It cannot write to the ledger and refuses to write content. On a 30-sentence seeded test that I wrote, it caught 20 of 20 unsupported claims, which is not independent proof. Its find-papers and find-gaps skills are untested, and it is still an unpublished draft.
Left: how much of my own review each screening route kept. Right: real verdicts from the Research Writer.
Key takeaway
The AI proposes. Code copies the quotes. A person checks.
Reflection
All four are prototypes tested on small or made-up samples, and none has been compared with human work, so none is proven. What I trust is the structure: the AI proposes, fixed rules limit it, and a person checks.
Grounded in
Code, test records and the list of claims I cannot make, for every build: the project repository on GitHub.
