From AI draft to publishing: what to hand over and what to check yourself

The line between what you can hand to AI and what you have to look after yourself is clearer than it seems.

Smoothing a sentence, or turning an idea into fuller prose, is fine to hand over. If you do not like the result you press generate again, and beyond a little time there is nothing to lose.

Deciding whether to send that piece out into the world under your own name is a different matter. Once it is published, correcting what went wrong costs far more.

---

Work you can hand to AI

Things you can undo easily at any point, and judge lightly by looking at the result.

• Writing the draft
• Cutting it down to each channel's character limit
• Converting the same content into a different tone
• Splitting a long piece into a thread

Every one of these can be thrown away the moment you look at it and do not like it.

---

Work a person has to judge

• Deciding which piece to post: Which of the proposed options suits your brand or account right now takes context to judge. What you covered recently, which topics are rising in your field, how readers will react, all of it has to be weighed together.
• Deciding when to post it: The date and time that suit a piece differ with its content and character.
• The final check before publishing: Easy to skip past, and the step most directly tied to how much your brand is trusted.

---

We applied that same line to the pre-publish checks

Tintio looks over a piece for typos and errors before it is published. Designing that, the question we spent longest on was not "what should we catch" but "what should we deliberately not treat as a problem". Here is the reasoning behind the five checks.

• Character count: Confirms whether a channel's limit has been passed. Platforms count characters differently, so the number in the editor and the number the check uses are built to run off one single standard. If the two disagree, you stop trusting the number on screen.
• Banned guide terms: Checks whether a term registered in your brand profile made it into the text. It does not reach out to context or synonyms; it catches exactly the terms you specified, without slippage. Broad drafting can go to AI, but the step that filters precisely has to run on a definite standard.
• Number cross-check: Compares whether a number that was not in the source material turned up in the generated result. It compares actual values, not their written form. If the source says 1,204 and the result says 1204, they are the same number. Where a thousands separator or a decimal point is ambiguous, the judgment is handled loosely, in the safe direction, so you do not get a needless warning. Numbers inside model or product names such as A4 or H100 are not flagged, and when no source material was supplied the check does not run at all, to prevent errors being thrown indiscriminately.
• Publishing language: Verifies in substance that the piece was written in the language set for that channel. Even when the AI reports "written in English", the text itself is analysed directly. Roman letters and Chinese characters are left out of scope, because English words and proper nouns mix naturally into Korean copy.
• Link validity: Confirms that a web address in the body actually opens. It reports only where a connection error was definitely confirmed, and in ambiguous situations such as a temporary delay it treats the link carefully rather than calling a healthy link a problem.

---

There is also a check we considered and dropped

At first we planned to catch a brand's more intricate writing rules automatically, along the lines of "when you use a number, put the basis for it in the same sentence".

During development we deliberately cut it. When a system enforces a rule without fully understanding the context, it raises warnings on healthy sentences and walks past the ones that actually need fixing. In the end the user only accumulates fatigue from correcting the tool's misfires one by one.

The heart of all five checks is one idea. Frequent false positives erode trust in the whole system. After three wrong error alerts in a row, the fourth one can point at something genuinely important and the user will no longer read warnings carefully. So we pick out only the definite errors and leave room for the judgments a person should make.

---

The check only points, the person edits

Tintio's checks do not let AI edit what they find. They tell you where something needs attention and then stop.

The moment a sentence changes without you knowing, it can become a piece quite unlike what you set out to write. A check alert is not a control line you must obey, so you can look at what was flagged and publish anyway. Deciding to look at the system's note and keep the text as it is, is also a meaningful human judgment.

---

The last decision always belongs to a person

To put it plainly: AI prepares the material and the drafts quickly and generously, and a person picks and approves the best of them.

From drafting through per-channel adaptation to the basic rule checks, the system handles it quickly. You look over the finished flow and make only the final decision about what goes to which channel and when. Not automating that last click is the core design principle Tintio holds to.

"If correcting a mistake costs more than time, a person makes the final check." That principle stays the safest standard no matter which automation tool you are working with.