What an AI-Native Workflow Looks Like — A Real Conversation and Three Ready-Made Templates

A real conversation ​

Open the reference design cooke_triplet.lens and say one sentence to the AI:

You: Change the thickness of the last surface to a paraxial focus solve, and see how much the total length changed.

AI: A dry run first; nothing has been written yet:

-surf 6    R=-18.39533   T=42.20778
+surf 6    R=-18.39533   T=solve(marginal_height,0)
-var R1..R6 T6
+var R1..R6

EFFL 50.0214 → 50.0214, BFL 42.4364 → 42.4364, FNO 5.00213 → 5.00213, TOTR 60.1767 → 60.4054. The focal length and F-number did not change, and the total length grew by about 0.23 mm: the original image plane was not at the paraxial focus (the back focal length is 42.4364, while the original thickness was 42.20778). Once the thickness becomes a solve, T6 is no longer an optimization variable. Shall I go ahead?

You: Go ahead.

These numbers were actually computed by LensBench; they are not illustrative. Note that the AI did not act right away: it first showed you which numbers would change and how the first-order quantities move, then waited for your nod.

Three ready-made workflow templates ​

Once connected to LensBench, three templates appear in the AI client's "prompts / slash commands" menu. Pick one and the AI works through fixed steps, each step following the same rules: changes are dry-run first, reported numbers carry an evidence id, and the verdict is whatever verify says.

1. Diagnose and optimize: survey (prescription, diagnostics, first-order quantities) → image-quality diagnosis (spot diagram, Seidel, MTF, field curvature) → confirm the targets with you and write them as specs → set the merit function and variables → optimize (can run in the background, with progress and cancellation) → verify → report with a before/after comparison
2. Zemax reconciliation: import the zmx → check the prescription surface by surface → compute first-order quantities, spot diagram, MTF, field curvature and Seidel with Zemax's sampling settings → list a deviation table → find the cause first, don't change the design first → export a zmx for a round-trip check if needed
3. Prepare report material: check the verdict with verify → export the pack (figures, numbers pages, tables, deterministic conclusions) → report; if the design changes, you are told the pack is stale

For template details, see Native AI.

When something goes wrong, hand that to the AI too ​

A crash, a wrong result, numbers that don't match Zemax: just tell the AI. It will help you export a feedback reproduction package, write the description for you following "what you did / what you expected / what actually happened" and let you look it over, and ask you first whether to attach the design file and screenshots. The package goes to the forum's "Bug reports" (ones with design files go through confidential feedback), where the forum bot reproduces it automatically and tracks its status.