Test Whether AI Follows Your Workflow's Instructions

Listen to this post

0:00 / 0:00

Industries

Tags

Content types

More time for your team.
Less time chasing tasks.

Connect the tools, tasks, and people behind your team’s daily work.

Instruction-following evals should test whether a model respects the constraints that make your workflow usable. A response that ignores the requested fields, invents a category, or performs an unapproved action can fail even when its prose sounds excellent.

Public benchmarks can suggest questions to ask. They cannot establish that your particular workflow follows its instructions.

Use historical benchmarks with context

Scale’s instruction-following evaluation page is marked deprecated as of January 2025. It describes a dataset of detailed prompts and human evaluation. Treat it as historical methodology, not a current model recommendation.

The useful lesson for a business team is to make instructions observable. “Be helpful” is difficult to score. “Return one permitted category and cite the request text supporting it” gives a reviewer something concrete to inspect.

Separate formatting from meaning

Some requirements can be checked mechanically: required fields, allowed values, length limits, or valid output structure. Other requirements need judgment: whether the answer addresses the request or whether a caveat changes its meaning.

Keep those checks separate. A structurally valid output can still be wrong, and a factually useful answer can still fail a strict interface contract.

For a request classifier, test both whether the category belongs to the allowed list and whether the chosen category matches the evidence. Include an explicit uncertain outcome for cases that do not support a confident choice.

Test conflicting and distracting inputs

Real inputs contain quoted emails, copied instructions, old information, and irrelevant text. Include examples where the source material asks the model to ignore the workflow’s rules.

The expected behavior should preserve the distinction between task instructions and material being processed. A customer’s pasted document should not gain authority to change recipients or access permissions.

Also test combinations of requirements. A model may follow a length limit and a source requirement separately while failing when both apply together. Your real interface probably combines several constraints.

Build a useful evaluation loop

  1. List the instructions that matter to the downstream process.
  2. Define a pass condition for each one.
  3. Create ordinary, ambiguous, and adversarial examples.
  4. Compare versions using the same cases and criteria.
  5. Inspect failure patterns before changing the prompt.

Keep critical boundaries outside the prompt where possible. If a tool must never send to an unapproved destination, enforce that restriction in the action path rather than relying entirely on a language instruction.

Test the workflow around Innflow

For an Innflow workflow, evaluate what happens after the model produces its output. Does invalid data enter review? Are unsupported categories rejected? Does a prepared action remain separate from execution?

Verify those behaviors in the actual implementation you build. A good model score does not establish a working connection, correct authorization, or successful downstream delivery.

Frequently asked questions about instruction following evals

Should we choose a model from an old leaderboard?

Use historical results for context, then test current candidates on your task. A deprecated ranking is not a current buying recommendation.

Can formatting checks replace human review?

No. They are useful for interface rules, but meaning and factual support often need additional evaluation.

What if a prompt change fixes one failure and creates another?

Review results across the case set. Keep the previous version available and evaluate the tradeoff before rollout.

Test the instructions that matter

Instruction-following evals become valuable when they protect a real handoff. Define the contract, test it, and explore Innflow for connecting model outputs to reviewable operational steps.

Related reading

AriannaKhan

Arianna Khan

Account Executive @ Innflow.ai

Continue

A clearer day starts with a connected flow.

Bring your team, context, and next steps together.

A little more, just for members.

Get Innflow updates and member offers by email.