Choose your AI assistant and control the budget

In brief

Use one coding assistant as your main tool and a repeatable review process. A chatbot can explain a traceback, edit a component or draft a test; it cannot certify your math or guarantee a three-star game. The cheapest useful setup is the one whose changes you can check.

• Start with tools you already have.
• Compare subscriptions with subscriptions, not with per-token API prices.
• Keep a spending ceiling and review the account's actual usage.

Price comparison: a dated snapshot

The figures below were checked on 6 September 2026. They are advertised US-dollar individual-plan prices, not a quote for Vietnam. Taxes, region, annual billing, promotions and account eligibility may change the checkout total. Usage limits differ, so equal prices do not buy equal development capacity.

Sources: OpenAI plan pricing, Claude plan comparison, and Cursor plan documentation. The workflow column is Slotgen's practical interpretation, not a measured ranking of these products.

Codex usage is included in eligible ChatGPT plans and has usage limits. API-key usage has separate usage-based pricing. Claude Pro includes Claude Code access but does not include Claude Console API usage. Cursor can incur additional on-demand usage charges. Read the Claude Pro explanation and each provider's current billing controls before enabling extra usage.

A budget that does not spiral

For a learning month, choose free access or one roughly $20/month plan if its limits meet your needs. Three $20 subscriptions total $60 before taxes and extras; they do not automatically triple progress. Save the rest for the bottleneck you actually encounter, such as original sound, animation work or an independent math review.

Keep four budget lines separate: recurring assistant subscriptions, metered API or cloud usage, software licenses, and commissioned work. Electricity and your own time are real costs even when a local simulator has no service fee. A local project folder does not mean every AI request is processed free on your computer.

Try the same small task in tools you can already access: explain a known failing test, change a label without altering state, and add an edge-case assertion. Record correct results, review time and usage. Choose based on that exercise rather than an unverified claim that one chatbot is always best.

Give the assistant an explicit boundary

A useful task prompt is specific enough to reject an attractive but wrong solution:

Inspect the current lines frontend and its package scripts first.
Change the bonus confirmation layout for narrow screens only.
Preserve mode IDs, bet amounts, RGS calls and event ordering.
Use the project's existing components and dependencies.
Show which files changed and run the available focused checks.
If a required API is unclear, locate its implementation;
do not invent a replacement or silently disable the feature.

For mathematics, ask for an explanation and tests before a new distribution. Supply the exact mode cost, payout units and a small known-good fixture. Ask the assistant to calculate the fixture independently. If it cannot explain why the expected number is correct, do not accept a larger generated model.

Separate authoring from verification

Use this loop: save a known-good checkpoint, request one bounded change, inspect the difference, run tests, replay the affected outcomes, then retain or revise the change. A second assistant can identify omissions, but agreement between two assistants is not proof. Both may reproduce the same mistaken assumption.

Keep a short project note containing the pinned SDK revisions, commands that actually work, event contract, asset naming rules and unresolved decisions. This reduces repeated explanation without filling the context with an entire art library or generated outcome book.

Never paste production session URLs, credentials, customer records or proprietary client assets into an unrelated service. Check privacy settings and the permissions of any connected tool. Give an assistant access only to the work it needs.

Use AI where the result is inspectable

Good early tasks include renaming an internal symbol consistently, writing a regression test, reviewing a small interface diff, explaining a build error and drafting a device checklist. Poor shortcuts include asking for a guaranteed-approval math model, accepting invented API endpoints, or shipping a pile of stylistically inconsistent generated images.

For art, AI can support internal exploration, but final assets still need intentional design, provenance and hands-on review. The visual production chapters explain how to build a coherent set without making the game feel like unrelated generated pictures.

Finish this part

Pick one assistant, set a monthly ceiling, and complete the small comparison exercise. Save one successful prompt and its verified result. Do not upgrade merely because a larger plan exists; upgrade only when a measured limitation is blocking useful work.