Action Items

Action Items: Extract Tasks from Any Text

Quickly turns call transcripts, chats, documents into a structured task list with assignees, deadlines, and priorities.

---

Step 1: Accept input

The user can provide:

| Format | How to process | |--------|---------------| | .txt file | Audio transcript (Whisper). Read entirely | | .json (Chat export) | Extract messages[].text, messages[].from | | .html (Chat export) | Extract message text from HTML | | .pdf | Extract text | | Multiple files | Process each, merge results | | Text in message | Process directly |

If no file provided, ask:

Attach a file (transcript, chat, document) or paste text
from which to extract tasks.

Handling large files

If file > 25000 tokens:

1. Read first 500 lines for context and participants
2. Process in blocks of 500 lines
3. Merge results, remove duplicates

Handling multiple files

If the user provides multiple files:

1. Process each in parallel via Task tool (subagent_type: "general-purpose")
2. Give each agent this same prompt for extraction
3. Merge and deduplicate results

---

Step 2: Identify participants

Extract all mentioned people/roles from the text:

• By name: "Alex said...", "Maria suggested..."
• By role: "client", "designer", "developer"
• User = "I", "me", "we"

Result: list of {name/role} for task assignment.

---

Step 3: Extract all commitments and tasks

Search the text for:

Explicit commitments

• "I will do...", "We will prepare...", "I will send by..."
• "Need to...", "Must...", "Should..."
• "Let's...", "Agreed that..."

Implicit tasks

• Questions without answers -> task "find out/answer"
• Problems without solutions -> task "resolve/investigate"
• Ideas/suggestions -> task "consider/evaluate"

Deadlines

• Explicit: "by Friday", "by February 15", "next week"
• Implicit: "urgently", "ASAP", "when there's time"
• Contextual: "before the meeting", "before launch"

---

Step 4: Classify and prioritize

For each task determine:

Priority:

• P0, Blocker: cannot move forward without this
• P1, Urgent: has a deadline or external pressure
• P2, Important: needs to be done, not on fire
• P3, Someday: idea, wish, backlog

Type:

• Decision (approve/decide)
• Action (do/create/send)
• Communication (write/call/discuss)
• Research (find out/verify/investigate)

---

Step 5: Generate report

Output format:

# Action Items: {source}
> Extracted from: {filename or description}
> Date: {today}
> Participants: {list}

## P0: Blockers
- [ ] **{task}** - {assignee} | {deadline}
  > Context: "{quote from text}"

## P1: Urgent
- [ ] **{task}** - {assignee} | {deadline}
  > Context: "{quote}"

## P2: Important
- [ ] **{task}** - {assignee}
  > Context: "{quote}"

## P3: Backlog
- [ ] {task}

---

## Unanswered Questions
- {question that was raised and not resolved}

## Key Decisions
- {what was decided during the conversation, for context}

---

Step 6: Suggest next step

After the report:

Extracted {N} tasks from {source}.

What's next?
1. Save to file (action-items-{date}.md)
2. Link to project via /session-save
3. Clarify/reprioritize

---

Rules

Accuracy over completeness

• Better to miss a questionable task than to invent a nonexistent one
• Every task must be backed by a quote from the source text
• If assignee unclear, write "unassigned"

Do not over-interpret

• Extract only what is said or clearly implied
• Do not invent tasks that are not in the text
• If context is ambiguous, note it

Audio transcript specifics

• Whisper may confuse names. Flag if a name seems uncertain
• "I guess we kinda should..." = weak task (P3)
• "Make sure you do this by..." = strong task (P1)
• Transcripts often contain false starts and repetitions. Filter them

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vasilievyakov
• Source: vasilievyakov/miracle-infrastructure
• License: MIT

Install and usage instructions live in the source repository linked above.