I would like to share a product proposal that emerged from repeated real-world use of ChatGPT rather than from an abstract feature idea.

The problem

In complex, multi-step tasks, ChatGPT can have enough context, reasoning capability and access to tools, yet still fail because the required specialized workflow is not explicitly enforced.

Across my own use, I documented 7 separate cases involving different task domains. The recurring failure modes included:

- incomplete recovery of prior context and related documents;

- missing preflight checks before acting;

- incorrect or incomplete tool/channel routing;

- loss of task state or pending actions;

- treating a previously identified issue as if it were new;

- insufficient verification before declaring work complete;

- excessive reliance on the user to remember what the system should have recovered itself.

One concrete example came from editorial review: after a high-effort ChatGPT review, a simple human pass still found a joined word (“continuardifícil”) and a substantive editorial/legal consistency issue. The important lesson was not that the model needed “more intelligence”; it needed a workflow that required the right checks before completion.

Proposal: Specialist Execution Contracts

My proposal is a versioned Specialist Profile / Specialist Execution Contract for ChatGPT.

A specialist profile would define, for a particular type of work:

- required state/context to recover before starting;

- mandatory preflight checks;

- tools that must or may be used;

- execution sequence;

- checkpoints and pending-action tracking;

- completion criteria;

- required evidence/verification;

- Evals or regression tests;

- versioning and rollback.

The goal is not to replace general reasoning. It is to make specialized execution explicit, reproducible, auditable and resumable.

Optional visual layer: persistent companions

The original idea started from ChatGPT’s visual companions/Pets.

A Pet could optionally represent a Specialist Profile and make its specialization, progress and verified workflow coverage visible to the user.

However, the Pet is only an interface layer. Competence should come from the execution contract, tools, verification and Evals—not from the character or gamification itself.

Why I think this is worth testing

The seven cases do not prove that this architecture is superior, nor do they establish population-level prevalence.

What they do show is a recurring product problem: having context, reasoning and tools available does not guarantee that the correct specialized sequence will actually be executed.

A useful next step would be a small controlled eval:

A: normal general-purpose workflow

B: the same task using a Specialist Execution Contract

Then compare:

- omissions / recall;

- false positives;

- user corrections required;

- duplicated actions;

- completion accuracy;

- time and cost;

- regression across versions.

Feedback I am looking for

I would be especially interested in feedback from people working with ChatGPT, Skills, Agents, Work, tool orchestration or Evals:

1. Does this problem framing match failures you have observed?

2. Could existing Skills/Agents/Evals already implement most of this pattern?

3. What would be the smallest useful MVP?

4. Would persistent structured task state be the critical missing component, or is there a simpler architecture?

I have a longer technical proposal and a structured evidence package covering the seven cases, but I am starting with this concise public version to invite technical discussion.