AGENCYSCRIPT
CoursesEnterpriseBlog
πŸ‘‘FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
Β© 2026 Agency Script, Inc.Β·
Privacy PolicyTerms of ServiceCertification AgreementSecurity

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

Treat the Image as a Cost Dial, Not a Free InputTile instead of supersizeAlways Tell the Model How to Weigh ModalitiesMake the Output Structured and TestableBuild an Adversarial Test Set Before You Trust AnythingVerify High-Stakes Fields in Code, Not in the ModelRedact First, Ask Questions LaterPick the Modality the Model Is Actually Good AtStart Narrow, Then Earn ScopeWhy this beats ambitionFrequently Asked QuestionsWhat is the single most impactful practice here?Do these practices apply to consumer use too, or just developers?Why structured output instead of natural language?How often should I re-run my adversarial test set?Key Takeaways
Home/Blog/Opinionated Rules for Your First Serious Multimodal Project
General

Opinionated Rules for Your First Serious Multimodal Project

A

Agency Script Editorial

Editorial Team

Β·April 23, 2026Β·8 min read
multimodal AImultimodal AI best practicesmultimodal AI guideai fundamentals

Most "best practices" articles for AI are a list of things everyone already agrees with: test your system, watch your costs, be careful with data. True, useless. The practices that actually matter for multimodal AI are more specific and more opinionated, and several of them contradict what you would do with a text-only model.

This is the list I would hand a team starting their first serious multimodal project. Each practice comes with the reasoning, because a rule you do not understand is a rule you will break the moment it is inconvenient. These are not theoretical. They come from the specific ways multimodal systems fail in production, the kind catalogued in 7 Common Mistakes with Multimodal AI (and How to Avoid Them).

Treat the Image as a Cost Dial, Not a Free Input

Text-only thinking treats inputs as cheap. With images, that mindset bankrupts you. Resolution drives both cost and latency, and the relationship is steep: a high-resolution image can cost as much as several pages of text.

The practice: resize every image down to the smallest dimensions that still pass your tests. Do not send the original. The reasoning is that beyond a certain point, more pixels add cost without adding accuracy, because the model downsamples anyway. You are paying to throw data away.

Tile instead of supersize

When detail genuinely matters across a large document, do not send one giant image. Split it into tiles at a readable resolution and combine the results. You get the detail without the cost explosion of one enormous image.

Always Tell the Model How to Weigh Modalities

This is the practice that separates people who have shipped from people who have only demoed. Models lean on text by default. If you do not tell them otherwise, they will quietly ignore what the image shows when it conflicts with the prompt.

The practice: in every multimodal prompt, state the precedence rule. "Base your answer on the image. If the user's description contradicts the image, trust the image and note the conflict." The reasoning is that you are correcting a known training bias, and the cost of leaving it uncorrected is silent, confident errors in exactly the cases that matter most.

Make the Output Structured and Testable

A free-text answer about an image is nearly impossible to evaluate at scale. You cannot diff prose reliably.

The practice: demand structured output, usually JSON with named fields, from the start. Instead of "describe the problem," ask for {issue, severity, location}. The reasoning is twofold. Structured output forces the model to commit to specific claims you can verify, and it gives you a target to grade against your test set. Vague output is unfalsifiable, which is the opposite of what you want in production.

The discipline behind this fits naturally into the workflow described in A Step-by-Step Approach to Multimodal AI, where the input-output contract comes first.

Build an Adversarial Test Set Before You Trust Anything

The happy-path demo is a trap. Your three clean test images do not represent the blurry, rotated, badly lit reality of user uploads.

The practice: assemble 20 to 50 cases that deliberately include the worst inputs, low resolution, conflicting image and text, rotated documents, empty images, alongside typical ones. Read every output by hand the first time. The reasoning is that you are buying information about failure modes before they cost you in production. Each surprising failure teaches you something you can design around now instead of debugging later under pressure.

Verify High-Stakes Fields in Code, Not in the Model

Asking a model to double-check itself helps, but it is not enough for anything that touches money, health, or law.

The practice: pull the model's structured output into code and verify what you can deterministically. If it extracted an invoice total, recompute it from the line items and flag mismatches. The reasoning is that a model's self-assessment shares the same blind spots as its original answer, while code-level checks catch the specific, costly errors, a wrong total, an impossible date, that humans most want caught.

Redact First, Ask Questions Later

Images and audio carry enormous amounts of incidental personal data that text rarely does: faces in the background, account numbers on screen, license plates, overheard conversations.

The practice: build redaction into the pipeline before any image or audio reaches a third-party model. Blur faces, mask sensitive fields, trim audio to the relevant clip. The reasoning is that the user uploaded a screenshot to ask one question, not to share everything else visible in it. Defaulting to redaction protects them and you.

Pick the Modality the Model Is Actually Good At

A spec sheet listing image, audio, and video support tells you what the model accepts, not what it does well.

The practice: verify each modality independently against your task before committing. Vision-language is usually the strongest; audio and video understanding often lag because paired training data is scarcer. The reasoning is that building your product on the model's weakest modality guarantees flakiness no prompt can fix. When in doubt, restructure the task to lean on the strong modality. The comparison in The Best Tools for Multimodal AI shows which tools are genuinely strong where.

Start Narrow, Then Earn Scope

The last practice is about restraint, and it ties the rest together. The strongest multimodal projects start with the smallest version of the task that delivers value, and expand only after the system has earned trust.

The practice: ship the lowest-stakes useful slice first. Instead of "read the invoice and post it to accounting," ship "read the invoice and suggest the fields for a human to confirm." The reasoning is that the failure cost of a suggestion is trivial, while the failure cost of an automated post is real money. A narrow first version lets you discover the model's actual limits on your actual data before those limits can hurt you.

Why this beats ambition

Ambitious first versions fail in expensive, visible ways that kill the project politically. Narrow first versions fail cheaply, teach you fast, and build the track record that justifies expanding scope. You expand because the data says you have earned it, not because the roadmap demanded it. Every practice above, resolution discipline, precedence prompts, structured output, adversarial testing, verification, gets easier to validate when the scope is small enough to fully understand.

Frequently Asked Questions

What is the single most impactful practice here?

Telling the model how to weigh modalities. It is one line in your prompt, and it corrects a training bias that silently corrupts exactly the high-value cases where image and text conflict. Skipping it produces confident, wrong answers that are hard to detect.

Do these practices apply to consumer use too, or just developers?

Most apply to everyone. Resizing and cropping images, being specific about what to look at, and verifying important answers help any user. The code-level verification and redaction pipeline are mainly for people building products.

Why structured output instead of natural language?

Because you cannot reliably evaluate or verify prose at scale. Structured fields force the model to commit to specific, checkable claims and give you a clear target to grade against. Natural language hides errors inside fluent sentences.

How often should I re-run my adversarial test set?

Every time you change the model, the prompt, or the preprocessing. Small changes can shift behavior in surprising ways, and the test set is how you catch regressions before users do. Treat it as a gate on every release.

Key Takeaways

  • Treat image resolution as a cost dial: resize to the minimum that passes your tests, and tile large documents.
  • Always state modality precedence in the prompt to correct the model's default text bias.
  • Demand structured, testable output so you can verify and grade at scale.
  • Build an adversarial test set early and verify high-stakes fields in code, not just in the model.
  • Redact images and audio before sending, and build on the modality the model is genuinely strong at.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way β€” a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026Β·11 min read
General

Case Study: Large Language Models in Practice

Most teams that fail with large language models don't fail because the technology doesn't work. They fail because they treat deployment as a one-time event rather than a discipline β€” pick a model, wri

A
Agency Script Editorial
June 1, 2026Β·11 min read
General

Thirty-Second Wins Breed False Confidence With LLMs

Working with large language models is deceptively easy to start and surprisingly hard to do well. You can get a useful output in thirty seconds, which creates a false confidence that compounds over ti

A
Agency Script Editorial
June 1, 2026Β·10 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification