Designing with the assistant
Asking for help
The assistant is good at the work you would do with a sequence viewer and a stack of references open. It is only as good as the constraints you give it.
What it can already see
You do not need to paste your plasmid into the chat. Within a project, the assistant can open your sequences, read the bases, run analyses against them and read your project notes. Say “the pET-28a construct” and it can go and look.
It also reads PLASMID.md at the start of every conversation in that project, which is where your standing constraints belong — project instructions.
What to say
Describe the experiment, not the software operation. “Insert this at position 1340” is a command you have already solved. “I need this expressed and purified from E. coli” lets it tell you that position 1340 is a bad idea.
The four things worth stating up front
- The host
- E. coli BL21(DE3), HEK293T, S. cerevisiae, a plant. It determines codon usage, promoter choice, selection marker and whether the protein can fold at all.
- What the protein has to do
- Purified from a cell lysate, secreted, localised to the nucleus, active as an enzyme, just fluorescent. A construct that expresses is not the same as a construct that works.
- What you have in the lab
- Your backbone, your enzymes, your antibiotics. A plan built around an enzyme you do not own is a plan you cannot run.
- What you cannot change
- A vector you must use, a tag your downstream assay needs, a site that must stay intact.
A request with enough in it
I need to purify the catalytic domain of my enzyme from E. coli BL21(DE3).
- Backbone: pET-28a(+), already in this project
- Tag: N-terminal His6 is fine, I purify on nickel
- The construct must be soluble — the full-length protein goes into
inclusion bodies, which is why I am cutting it down
- Enzymes I have: NdeI, XhoI, BamHI, EcoRI, HindIII
Plan the cloning and tell me if the domain boundaries look wrong.That is not a long message, and it removes almost every assumption the assistant would otherwise have to make on your behalf.
Pointing at something specific
- A region
- Select it on the map, right-click, and choose Send to AI Assistant. The selection travels with your question, so “what is this?” is a complete question.
- A file
- Attach it in the composer — a sequence file, a spreadsheet, a PDF, or a photo of a gel.
- A construct by name
- Just name it. It will find it in the project.
When it asks you something
Sometimes it will stop and ask before continuing — which host, which marker, which of two designs. These are the decisions where guessing would waste your time, so answering carefully is worth more than it looks.
Answers appear as options in the composer. If none of them is right, type your own — free text is accepted and is treated as your answer, not as a stray comment. Your answer becomes a constraint the rest of the work is held to.
When an answer looks wrong
Say so, and say why. Two things that work much better than starting over:
- Correct the premise, not the output
- “You have used XhoI but my construct has two XhoI sites” gives it something to act on. “That is wrong” does not.
- Ask it to prove the specific claim
- “Show me the translation from the start codon through the junction” turns an assertion into something you can check yourself. Asking for the evidence is always fair, and it is the fastest way to find out whether an answer is real.
Before it changes anything
For anything beyond a small edit, switch to Plan mode first. It cannot write to your files there, so you get to read the strategy before any DNA moves. Plan and Act explains the difference, and approvals covers where it stops to ask.