Adopt AI in your business through a practical first project
Choose a first business AI project, prepare data and responsibilities, then compare accepted outcomes before expanding AI adoption across your team.
8 min read
Reaching a goal with AI sometimes means building the tool that lets you get there. An adjustable preview, a comparison tool or a small test harness can unblock work that repeated prompts keep circling around. ParamRig applies this idea to visual systems: the agent prepares the controls, and you tune the result.
The detour earns its place when it makes the next decision easier. You should be able to observe something, try a variation and keep a precise choice. If the tool does not help you do that, return to the intended outcome.
Notice when a tool is missing
You ask for a smoother animation. The agent changes its duration. It now feels too slow. You explain that the opening of the movement was the real issue. Another version improves the acceleration but changes the ending as well. After several exchanges, it is difficult to compare what actually changed.
The way you express the choice may be the problem. “Smoother”, “slightly less dense” and “more balanced” describe a perception. Choosing an exact value calls for variants you can see and manipulate.
Other obstacles need other tools. A data importer might lack a preview of rejected rows. A research task might lack a comparison protocol. An agent fixing a bug might lack a reproducible case. Producing more versions of the final result does not necessarily fill that gap.
Ask one short question: what must I be able to see, measure or adjust to choose the next step? The answer defines the next tool’s scope.
Take the smallest useful detour
A project tool does not have to become a complete application. It might be one page with a preview and a few controls, a script that compares two outputs, or a reproducible test scenario.
Scroll to read all columns
| Obstacle | Tool to build | Decision it should enable |
|---|---|---|
| A visual adjustment is difficult to describe | Preview with exposed parameters | Choose values while looking at the result |
| Two approaches appear equivalent | Comparison on the same input | See the differences that affect the choice |
| A defect keeps returning | Minimal reproducible case | Trigger the defect and verify a correction |
| Data is hard to check | Transformation preview | Catch an error before applying the whole batch |
Keep the same input during comparisons. If the agent changes the content, layout and parameters together, you cannot tell what explains your preference. A stable reference often helps more than another variation.
Define an exit condition, too: for example, being able to choose and retrieve an animation’s values. Once that condition is met, use the tool to move the project forward.
ParamRig: ask the agent to build the controls
ParamRig is an open-source workbench for tuning visual systems built with coding agents. Its workflow guide describes a simple loop: the agent exposes useful parameters, you adjust them in a preview, and the agent receives the values you kept.
A set of controls, called a rig, can surround a component, SVG, animation, 3D scene or page. Controls can represent numbers, colors, curves or choices. You manipulate the result rather than turning every adjustment into another sentence.
ParamRig supports reference comparison, snapshots and animating values on a timeline. Depending on the path used, the handoff is an export or a reviewed batch of changes in the connected project. The agent receives a concrete choice: these are the values to keep and what they should change.
ParamRig is another project in our ecosystem. We discuss it here because this division of work addresses the problem directly: the agent builds the means to explore, and the person makes the creative decisions.
Example: finding the right movement for a 3D scene
Imagine a scene presenting a product. Its geometry works, but the rotation and lighting do not yet have the right rhythm. Asking the agent to “make the whole thing more elegant” leaves many choices unstated.
Goal: a readable movement that shows the product without distracting from its shape.
Tool requested: a preview of the same scene exposing rotation speed, camera angle, light intensity and the animation curve.
Exploration: preserve the reference scene, change one setting at a time and keep the variants that meet the need.
Expected output: selected values, their reference and instructions for applying them in the project.
Verification: replay the scene in its final context to confirm the movement and readability.
The tool makes the choice more precise. It does not decide what suits the product for you. Final verification still matters: a movement that feels good in the workbench might take too much space or run too long once integrated into the page.
The same reasoning applies to an animation curve, a component’s proportions or a vector composition. Expose only the parameters that affect your decision. A panel full of irrelevant controls moves the difficulty elsewhere.
Give an AI agent the tooling task
Start by explaining the outcome that is blocked, then the decision the tool must let you make. Ask the agent to preserve a reference and make choices recoverable. A project brief remains useful for temporary tooling.
Adapt this prompt to build the missing tool
Project goal: [intended outcome].
Current obstacle: [what I cannot yet observe, compare or adjust].
Propose the smallest tool that lets me make this decision. Explain what I will be able to verify with it.
For a visual result, examine whether a ParamRig rig can expose the useful parameters around a preview.
Preserve a stable reference. Include only controls that affect the choice and explain their effect.
Make the selected values retrievable so they can be handed back for integration into the project.
Tool exit condition: [decision made possible]. Final verification: [outcome to check after integration].
If the obstacle is still unclear, request a short exploration before construction. The agent should distinguish missing information, an error and missing controls. Building a settings panel will not repair an incorrect data source.
Organize the detour in Stellary
In Stellary, keep the original goal visible and link the tool-building task to it. The tool’s card describes the capability you need: “compare curves and preserve selected values”. The outcome card describes the delivery: “integrate and validate the animation in the product page”.
A workflow with a supervisor can divide this work: an explorer identifies the gap, a creator builds the tool, the person selects settings, an executor applies them, and a reviewer checks the result. Use the roles the work needs; a simple task can stay with one person and one agent.
Name the blocked decision
State what is missing before you can choose the next step.
Build the means to decide
Create the preview, comparison or test that addresses the gap.
Explore and preserve the choice
Compare against a reference and keep usable values or evidence.
Return to the outcome
Integrate the choice into the project, then check the final user flow.
Keep the link between the tool, the choice and the integrated result. It explains why values were selected and helps avoid repeating the exploration during the next change. This record can join the project documentation.
Know when to stop building tools
A detour becomes expensive when the tool takes the goal’s place. Adding a preset library, account system or new editor does not necessarily help you choose the curve blocking delivery.
For each addition, ask which decision it enables. If there is no answer, save the idea for later. A one-off problem may only need a disposable tool. If the same need returns, you can then make it easier to reuse.
For your next visual obstacle, start with a ParamRig example or ask your agent for a preview with a few controls. The first outcome to seek is simple: being able to make the decision that kept the project from moving forward.
Connect the tool to the project outcome
In a supervisor workflow, give tooling a stopping condition: compare two variants, preserve a setting or reproduce a defect. Stellary’s coordination setup keeps that task attached to the deliverable it should unblock. The visual example illustrates the method; it does not measure a productivity gain.
