At iO, we use an internal AI tool to generate first-draft Jira tickets from a feature description. Scope, acceptance criteria, edge cases, analytics requirements. Sounds like a small thing — but a well-structured ticket is excellent context for Cursor, and that changes the entire development workflow.
The ticket generation step
A good AI-generated ticket has:
- Scope: what's in, what's explicitly out
- Acceptance criteria: specific, testable conditions
- Edge cases: the scenarios that are easy to miss
- Analytics: what events to track, what properties to include
These are generated from a brief description, not written from scratch. The model fills in the structure and forces you to think about edge cases you might have skipped.
Using the ticket as Cursor context
Once you have a well-structured ticket, paste it into your Cursor session, add the relevant code files, and the model has everything it needs.
Acceptance criteria become test cases. Edge cases become guardrails. Scope definition prevents the model from going off on tangents.
For UI tasks, grab the Figma link for the component you're implementing. In Figma, select the component or frame you want to build, right-click and choose Copy/paste → Copy link to selection (or press ⌘L). Add it to your Cursor prompt as:
Figma design: https://www.figma.com/design/abc123/...?node-id=12-34
Install Figma's Cursor integration from Customize → Marketplace, then authenticate once. You can also configure the remote MCP server manually if your team manages MCP connections outside the marketplace. Follow the current Figma MCP setup guide because the exact installation surface can change between Cursor releases.
Setting up the Figma MCP
Figma MCP gives Cursor direct access to design tokens, component variants, and layout data. The remote server is the recommended setup; the desktop server remains available for organization-specific local workflows. See Figma MCP Workflows for setup.
The flow: ticket context + Figma design link + Figma MCP → plan → implementation → PR. The PR description writes itself from the ticket's acceptance criteria.
Model strategy
Ticket generation benefits from a capable model — you want it to think carefully about edge cases. Implementation can use a faster model once the context is set. Final review? A reasoning model.
Model Mixer
Assign models to steps and compare workflow cost assumptions
Pick a task
Review a medium-sized file and get structured feedback on issues and improvements.
~4k in / ~1.5k outHow often?
600 runs / 30-day month · 4,000 input + 1,500 output tokens per run
GPT-6 Luna is 99% cheaper than Claude Fable 5 at these token assumptions — $0.690 vs $69.00/mo
A fast model catches obvious issues. A stronger model finds subtle logic bugs and gives more actionable feedback — fewer missed problems means less rework later.
Full registry checked 2026-10-05; Mistral checked separately 2026-10-07. Monthly estimates hold the current rate constant. Mistral · Anthropic · OpenAI · Google · DeepSeek · Moonshot AI · Z.ai · Cursor
Pricing assumptions and capacity limits
Standard uncached text-token estimates. Excludes cache writes/reads, tool fees, taxes, and platform-specific charges. Reasoning tokens count as output. Mistral Large 4 checked separately on 2026-10-07: public preview, $1.36/$4.18 standard input/output; $0.68/$2.09 launch rates. The two-week offer is interpreted as October 6–19 inclusive. Other models retain the full-registry verification date above. Claude Opus 5.5 is $4/$20 and Sonnet 5.5 is $2/$10. Opus 5 and Opus 4.8 remain available at $5/$25 and appear only in historical examples. Opus 5.5 fast mode is $8/$40. GPT-6.1 Sol is $2/$10 and GPT-6 Luna is $0.10/$0.50. GPT-5.6 Luna, Terra, and Sol remain $0.20/$1.20, $2/$12, and $4/$20; OpenAI calls the Sol rate promotional through at least November 21, 2026. GLM-5.3 uses Z.ai API list rates ($1.40/$4.40). GLM-5.3-Flash ($0.15/$0.50) is not in the calculators. Gemini 3.8 Flash is $0.75/$3.75 through December 31, 2026, then $1.50/$7.50. Google API rates are used; Cursor currently lists a different output rate. DeepSeek-V4.1-Flash estimates use peak uncached rates ($0.30/$1.20). Off-peak rates are half. Retired V4-Flash rates are historical only. OpenAI requests above 272K input tokens use 2× input and 1.5× output rates. Gemini 3.1 Pro applies those multipliers above 200K. Context windows describe provider API capacity, not a universal Cursor default. Availability and limits vary by platform. These are cost estimates, not a capacity check. Verify context and output limits for large requests.
1. Scaffold
Component structure, boilerplate, file setup
2. Business logic
Core feature logic, edge cases, state
3. Code review
Architecture, patterns, security concerns
4. Tests
Unit tests, edge case coverage
Estimated cost per completed workflow
Each attempt repeats the full token budget. Model choice does not guarantee first-attempt success. Adjust both assumptions to reflect your experience.
Provider API estimates. Full registry checked 2026-10-05; Mistral checked separately 2026-10-07. Mistral’s 50% launch offer is interpreted as October 6–19 inclusive. Excludes subscription allowances, caching, and tool fees.
Workflow Recipe
Copy-pasteable flows and guided workflow finder
Pick a workflow
Prompt
I need to build [feature]. Here’s the spec: inputs, outputs, edge cases, constraints. Produce a plan before writing any code.
Steps
Spec
Write a developer spec with inputs, outputs, edge cases, and constraints.
Plan
Ask the model to produce a plan — files to create/modify, key decisions. Review before coding.
Code
Implement against the plan. One feature, one PR. Pull the model back if it goes out of scope.
Review
Ask the model to review its own code: "What edge cases might this miss? What would break this?"
Tests
Generate tests from the review’s edge cases: "Write tests for the edge cases you identified."
PR description
Generate the PR description from the spec and the diff. Full context produces clear descriptions.
Tools
Guardrails
- One feature, one PR — keep scope tight
- Review the plan before writing code
- Don’t let the model touch files outside scope
- Ask before refactoring adjacent code
Expected output
Working PR with passing CI, clear description, and tests covering the identified edge cases.
Paste into CLAUDE.md, .cursor/rules/, or your workflow docs
What are you trying to do?
What are you trying to do?