Parse Reply

Input

Text of the user's reply to a brief email.

Pre-processing

1. Strip HTML (if input is HTML)

If the input contains HTML tags, extract plain text before doing anything else:

• Remove all HTML tags (e.g. <div>, <p>, <br>, <span>, <b>, <blockquote>, etc.)
• Decode HTML entities (  → space, & → &, < → <, > → >, " → ")
• Collapse multiple consecutive blank lines into a single blank line
• Strip leading/trailing whitespace from each line

2. Strip quoted original email

After HTML stripping (or if input was already plain text), remove everything below the first occurrence of any of these separator patterns:

• A line starting with From:
• A line starting with Sent:
• A line that is exactly --- or ___ or ***
• A line starting with > followed by On ... wrote:
• A line starting with ________________________________ (Outlook's horizontal rule in plain-text format)
• A line matching On .{5,80} wrote: (Outlook's "On Mon 24 Mar, Dayarc <...> wrote:" pattern)

Only parse the text above the first separator.

3. Strip email signatures

After removing quoted text, also discard lines from the first occurrence of a signature block marker:

• A line that is exactly -- (standard signature delimiter)
• A line containing only Sent from my iPhone / Sent from my Android / Get Outlook for iOS (and similar auto-appended footers)

Only parse the text above the first signature marker.

Output (JSON)

{
  "corrections": [{
    "action": "mark_done",
    "target": "auth migration",
    "detail": "User says auth migration is complete"
  }],
  "quality_signals": [{
    "sentiment": "negative",
    "detail": "too much noise in inbox section"
  }]
}

Instructions

The user may write in any free-form style — there are no required keywords or formats. Read each sentence for intent and extract meaning, even when phrasing is indirect or implicit.

1. Parse natural language corrections. Infer the action from context — do not require specific trigger words:
   • mark_done: any sentence conveying that something was completed, sent, fixed, resolved, or handled. Examples of intent (not exhaustive): "X is done", "finished X", "I already fixed the issue related to X", "I already sent the mail to my peer about X", "that's shipped", "merged", "resolved the X problem".
   • remove: any sentence conveying that something is no longer relevant, was cancelled, doesn't need tracking, or should be dropped. Examples: "drop X", "X got cancelled", "no longer relevant", "ignore X", "we decided not to do X".
   • add_priority: any sentence conveying a new thing to track, do, or remember. Examples: "remind me to X", "I need to X", "add X as priority", "don't forget X", "new priority: X", "I should chat with PM".
   • correct: any sentence conveying that existing information is wrong or needs updating. Examples: "actually X is ...", "correction: ...", "X should be Y not Z".
   • For indirect references (e.g. "I already fixed the issue related to the UI"): extract the most specific identifiable subject as target (e.g. "UI issue") and note the indirect phrasing in detail.
2. Parse quality signals — sentences that evaluate the brief itself (not specific work items). Classify sentiment and extract detail:
   • positive: any expression of approval, usefulness, or accuracy about the brief. Examples: "great brief", "spot on", "exactly right", "really useful", "perfect", "loved the summary".
   • negative: any expression that the brief was noisy, inaccurate, incomplete, or off-focus. Examples: "too much noise", "priorities were off", "missed my X work", "not relevant", "wrong focus", "too long", "cluttered".
   • Map each match to { sentiment: "positive"|"negative", detail: "<normalized phrase or quoted fragment>" }.
3. Ignore purely social/phatic replies with no work content: "thanks", "looks good", "ok", "got it", "👍". These produce neither a correction nor a quality signal.
4. A reply may contain both corrections and quality signals — extract all.
5. If no actionable content found, return empty arrays for both corrections and quality_signals.