AI Prompts to Find Deadlines and Commitments in Email

The missed deadline buried in email rarely looks dramatic when it happens. A thread runs to twelve replies over three days, someone writes 'I'll get that across to you by end of the week,' and end of the week passes quietly. Nobody followed up because both parties read the same thread and each assumed the other owned the action. That is the typical shape of a commitment that gets dropped: ambiguous, buried, and never acknowledged as a deadline in the first place.

AI is well-suited to this kind of extraction because it reads every line at the same level of attention, without skimming the middle and without treating a casual sentence as background noise. The problem is that most people prompt it wrong. Asking for a 'summary' returns narrative prose; what you need is a structured list where each row is one obligation, with columns for who made it, exactly what was promised, and whether the date is firm or vague.

The three prompts below target different categories of obligation: what you committed to, what others committed to you, and hard external deadlines that exist independent of anyone's word. Each includes a rule that most guides skip — flag anything phrased vaguely rather than converting it silently to a date. That distinction between 'Friday at 5 pm' and 'sometime early next week' is precisely what disappears in a generic summary and precisely what causes a dropped commitment six days later.

What do you need before you run these prompts?#

Prepare the thread before pasting it. Open the conversation in your email client and expand the full quoted history — the 'show trimmed content' control in Gmail, or the ellipsis that reveals quoted blocks in other clients. If you copy only the visible top message, the model sees one reply and misses every commitment made in the chain below it.

Keep the sender names and dates in the paste. A commitment without an owner is not actionable: knowing that someone promised to send a document is less useful than knowing Maria promised it on Tuesday. Paste the thread oldest message first so the model reads the conversation in the order it happened, which matters when a deadline was moved and you want to know which date is current rather than which appeared first.

Strip the noise before pasting: signatures, legal disclaimers, and repeated quoted blocks add length without adding facts. In a long thread they can also confuse the model, which may read a confidentiality footer as content. A quick pass to remove obvious boilerplate takes thirty seconds and consistently sharpens the output.

How do the three prompts work?#

Each prompt targets one category so the outputs stay clean and easy to act on. Running them in sequence — your commitments first, then what others owe you, then external deadlines — keeps each list short and focused. You can run all three in the same conversation: after the first output, paste the next prompt and the model carries the thread context forward without you re-pasting the thread.

1. 1

   Prompt 1: find what you personally promised

   Ask only for commitments you made. The 'exact phrase used' column forces the model to quote rather than paraphrase, which is the main protection against an invented date appearing in your obligations list.

2. 2

   Prompt 2: find what others promised you

   Switch the perspective and ask for commitments others made to you. The vague-flag rule applies here too: 'I'll look into it' should appear flagged rather than converted into a firm due date your tracker will then treat as real.

3. 3

   Prompt 3: sweep for hard external deadlines

   External deadlines differ from personal commitments: a contract date or regulatory cutoff exists regardless of what anyone in the thread said. Separating them produces two distinct lists — one about trust between people, one about facts that will not move because someone asked nicely.

How do you find every commitment you personally made?#

The key instruction in this prompt is the exact-phrase column. Asking the model to quote the words you used rather than paraphrase them closes the gap through which most hallucinated deadlines enter. If the model has to reproduce what you wrote, it cannot add a date you never stated.

The vague flag is equally important. Without it, a model will often convert 'I'll try to get this across to you early next week' into a Monday entry on the obligations list. That output looks authoritative and is misleading. Demanding the [VAGUE] tag forces an honest representation of what was actually committed and what was only intended.

How do you find what others promised you?#

The perspective flip is the only structural change from Prompt 1. Every rule carries over and the logic is the same: an exact quote prevents invention, and the vague flag prevents a soft 'I'll look into it' from appearing as a firm deadline that you then follow up on unnecessarily — or worse, that you let slide because it looked too vague to track.

One additional rule is worth adding for long or contentious threads: ask the model to reflect the latest state only. If someone committed to sending a report on Wednesday and then, three replies later, pushed it to Friday, you want Friday on the list, with a note that it moved, not both dates with no indication of which is current.

How do you sweep for hard external deadlines?#

External deadlines — contract dates, regulatory windows, client-set cutoffs, event dates — behave differently from personal commitments because they do not move when someone says they will try harder. This prompt isolates them so they sit on a separate list from the things people promised each other, which makes it clear at a glance what is negotiable and what is not.

The 'flag if it moved' rule is worth including here because external deadlines do sometimes shift, and a thread may reference both the original and the revised date without clearly distinguishing them. Knowing a date changed once already, before you act on it, is useful context.

How do these results vary across AI tools?#

The three prompts work across the major chatbots, but the outputs differ in predictable ways. The table below reflects documented model behavior as of mid-2026; verify against the current version of any tool you rely on, since model updates change these characteristics.

What do you do when the output is wrong?#

Most extraction errors trace to one of four causes. In three of the four, the fix is in the input rather than the prompt.

• Thread pasted incomplete or out of order: the model summarizes one message and presents it as the whole thread, or reads the newest reply as the opening context. Fix: expand the full history before copying, paste oldest-first, and confirm you have the complete conversation before running any prompt.
• Commitment was implied, not stated: someone replied 'that sounds good' to a proposed deadline without explicitly agreeing. The model may flag this correctly as ambiguous, or it may log it as a firm commitment. When the stakes are high, verify any extracted obligation against the original message.
• Model invented a date for a vague promise: the hallucinated deadline looks identical to a real one because it is stated with the same confidence. The 'quote exactly' rule reduces this risk significantly, but if a date appears in the output that you do not recognize from the thread, search for it in the source before acting.
• Thread too long for a reliable single pass: quality degrades on very long threads because the important sentences are diluted. Split by date range, run each chunk separately, and combine the outputs by hand — the same chunking approach that works for summaries works here.

Is there a faster way to track commitments without pasting?#

The copy-paste loop works, but it is manual: open the thread, expand the history, copy, switch tabs, paste, run the prompt, copy the output, and decide where to file it. For one thread that is a reasonable trade. For ten threads every morning it is a process that gets skipped under pressure, which means the threads that most need the sweep — the long, complicated ones with the most at stake — are the ones it does not happen on.

AI Emaily is an AI-native email client that extracts commitments, deadlines, and action items inside your inbox without a paste step. The model reads the full thread where it already lives, with the sender names, the dates, and the quoted history intact and in order, and surfaces what you owe and what others owe you alongside the conversation. We build AI Emaily, and the obligation extraction it runs is the same job these prompts do manually — continuously, on every thread, without a tab switch. You can start a 7-day free trial at aiemaily.com, or review plan pricing at aiemaily.com/pricing before signing up.