Pre-spend planning for agent-generated media

Recover an overbroad request before it becomes paid trial and error

If the request is simply “make campaign media,” stop before generation. Rewrite it as one deliverable, one intended placement, one audience action, one visual premise, and one reviewer. Then record the source-rights status, prohibited claims, required subject details, editable text needs, ratio, and pass condition. A narrow brief makes a failed proof diagnosable; a mixed brief can spend credits without revealing whether the model, prompt, frame, or source input caused the miss.

Run the live catalog and balance checks again whenever the plan changes. The published MCP package separates `list_models`, `get_credits`, and `generate`, so model discovery and balance inspection can happen before the paid call. Its `generate` schema requires a model identifier and prompt, with optional ratio and wait controls. Do not reuse a remembered catalog cost or assume that a model selected for a still can accept an image-to-video input.

When a proof misses, classify the miss before retrying: wrong workflow, wrong model, ambiguous prompt, unsuitable ratio, unresolved source image, or an editorial requirement the generation schema does not control. Revise only one material variable for the next proof. If the balance cannot fund that bounded retry plus the rest of the approved plan, stop and re-scope instead of turning every rejected output into another automatic call.

Brief gate

One deliverable and one pass conditionName the placement, audience action, source constraints, reviewer, and stopping rule.

Catalog gate

Current model ID, workflow, ratio, and creditsRead live model data and the current balance immediately before approval.

Retry gate

One diagnosed changeDo not buy a larger batch to hide an unresolved first-proof failure.

Keep the approval record separate from the agent conversation

A useful pre-spend record is small and explicit: request owner, approver, model ID, workflow, displayed credit cost, current balance, prompt version, ratio, source input, maximum call count, and stop condition. Store that record where the production team can review it. Natural-language agreement in a long agent thread is not a durable budget, and the current package does not expose a separate MCP cost-estimation or payment tool.

The package documentation says top-up discovery and Checkout creation are CLI or API features rather than MCP tools, and that those paths return a URL for a human browser action instead of charging a payment method directly. That product boundary supports a human funding step, but it does not itself enforce a campaign budget. The operator still needs to compare the approved proof plan with the balance returned by `get_credits` and limit how many generation calls the agent may request.

Treat refund paths narrowly. OfflineCreator currently says failed jobs return reserved credits automatically. The version-pinned server describes `cancel_generation` as cancelling a reserved generation before provider submission and refunding credits. Neither statement makes an editorially rejected completed output free, and neither justifies assuming that any submitted job can still be cancelled. Record the actual generation state and observed balance change rather than forecasting a refund as available budget.

Use the current credit table to choose the smallest informative proof

The August 2026 first-party catalog lists six launch models across three workflows. Text-to-image costs are 1 credit for FLUX Schnell, 5 for Recraft V3, and 8 for FLUX 1.1 Pro Ultra. Kling 2.6 Pro text-to-video and Kling 2.6 Pro Motion image-to-video are each listed at 42 credits. Veo 3.1 Fast text-to-video is listed at 96 credits. These are current per-generation catalog charges, not quality scores, supplier guarantees, or a quote for a finished campaign.

Choose by the uncertainty the proof must resolve. A 1-credit FLUX Schnell call can test whether a rough still direction is worth refining. Recraft V3 or FLUX 1.1 Pro Ultra can test a more specific still route after the visual premise is approved. A 42-credit motion call should answer a motion question that a still cannot. A 96-credit Veo call should not be the first place the team discovers that the brief, ratio, or campaign premise was unclear.

Before approval, calculate both the planned total and the worst permitted retry total using the live rows. Keep those arithmetic scenarios outside the prompt so a reviewer can see them. The cost table may change, and the package's `list_models` tool is expressly described as returning the launch catalog with credit costs, so the live result should override this editorial snapshot at execution time.

Rough still direction

FLUX Schnell · 1 creditUse only when a low-cost composition or atmosphere check can answer the current question.

Brand or detailed still proof

Recraft V3 · 5 credits or FLUX 1.1 Pro Ultra · 8 creditsChoose from the catalog's stated fit; do not present the labels as comparative test results.

Social or image-led motion proof

Kling route · 42 creditsApprove motion purpose and source-image status before making the paid request.

Premium text-to-video proof

Veo 3.1 Fast · 96 creditsEscalate only after cheaper planning checks have resolved the brief and frame.

Make the proof budget visible with scenario arithmetic

Suppose a team wants a square brand still and one short social-motion concept. At the current catalog rates, one Recraft V3 proof plus one Kling 2.6 Pro proof is 47 credits. Allowing exactly one approved retry of each makes the ceiling 94 credits. That calculation is transparent because it multiplies the listed per-generation charges; it does not include downstream design, editing, storage, taxes, or any promise that both outputs will pass review.

A lower-cost discovery sequence could use one 1-credit FLUX Schnell draft before a 5-credit Recraft V3 still, then stop before motion unless the still establishes the visual premise. Adding one 42-credit Kling motion proof makes that staged plan 48 credits. The Studio template catalog similarly presents separate 1-credit workspace mood, 5-credit geometric poster, 42-credit social hook, and 96-credit workshop glide recipes. Separate examples reinforce that each call should have its own question and approval.

Do not turn scenario arithmetic into an instruction to exhaust the ceiling. The ceiling is permission to make at most the named calls, not a target. After every output, compare the result with the pass condition and update the remaining balance from the product. Stop when the deliverable is answered, when a required input remains unresolved, or when the next call would exceed the approved ceiling.

Still plus motion

5 + 42 = 47 creditsOne Recraft V3 still proof and one Kling 2.6 Pro motion proof at the current listed rates.

Same plan with one retry each

47 × 2 = 94 credits maximumA reviewer-approved ceiling, not a recommendation to spend every allowed credit.

Staged discovery

1 + 5 + 42 = 48 creditsAdvance from rough still to refined still to motion only when the preceding gate passes.

Choose this workflow when uncertainty can be reduced before generation

Pre-spend planning fits a team that already knows the intended deliverable and can pause between model discovery, approval, generation, retrieval, and review. It is especially useful when a cheap still can resolve composition before a motion request, when image-to-video requires a rights-cleared source, or when several stakeholders need to agree on a hard call ceiling.

Avoid an automated path when the requester cannot name a pass condition, the input rights are unresolved, the account balance is unknown, or the job depends on capabilities not present in the request schema. The current `generate` contract accepts model ID, prompt, optional 16:9, 9:16, or 1:1 ratio, and optional waiting. Image-to-video uses a reserved generation followed by `upload_input`. Those controls do not replace brand review, factual review, source authorization, downstream editing, or destination-specific export checks.

The decision is not “which model is best?” in the abstract. It is “which smallest current call can answer the next unresolved production question?” If no listed call can answer that question without unsupported assumptions, spend zero credits and fix the workflow first.

Continue from the next unresolved decision

Return to the creator-use-case hub when the output type or production cadence is still undecided. Use deterministic finishing when the generation plan is clear but editable typography, layout, captions, or export rules remain. Use draft-first publishing when the media proof is complete and the next risk is an unreviewed handoff to a public channel. These routes preserve this page's narrow ownership of the budget and approval decision before generation.

Keep claims inside the pre-spend evidence boundary

This page owns “ai generation pre spend planning”: defining the brief, checking rights and inputs, reading the current model and credit rows, inspecting balance, approving one call ceiling, and stopping between proofs. It does not own a general MCP use-case directory, model-quality ranking, complete campaign budget, supplier-cost forecast, creative benchmark, or publishing workflow.

The last30days v3.18.4 run returned 92 recent-source records, but every evidence cluster scored zero and Reddit ended partial after RedditRSS feed failures. That gap is not evidence that practitioners are silent; it means this draft cannot claim community consensus, adoption, savings, reliability, output quality, conversion, or customer outcomes. Current model and tool statements therefore come from retrieved first-party product pages and a version-pinned public package, with catalog labels explicitly treated as vendor positioning rather than independent evaluation.

Refresh model names and credit values before using this page as an execution checklist. Preserve draft status until editorial review confirms that the arithmetic, product surface, and linked next actions remain current. If current model and credit data can no longer support a distinct pre-spend decision path, consolidate the topic into the parent creator-use-case hub.