AI Transcript Action Extractor

intent

this skill converts pasted transcripts, chat exports, meeting notes, or discussion threads into a clean, reviewable action register. it extracts concrete tasks, owners, deadlines, confirmed decisions, risks, dependencies, and open questions from raw text the user provides. use it when you need to turn messy discussion into a structured work plan without fabricating details or accessing external systems. the skill works only with text the user pastes directly and includes built-in safeguards for redacting sensitive data before external AI use.

inputs

• transcript or chat export: raw pasted text from a meeting, chat log, support thread, interview, standup, or discussion (required)
• meeting or discussion context: brief description of what the conversation was about (optional but improves accuracy)
• participant names and roles: list of who spoke and their positions, if known (optional)
• date of discussion: when the conversation happened (optional)
• desired output format: preference for table, bullets, CSV rows, or recap-ready register (optional, defaults to table)
• sensitivity flags: whether the text contains confidential, personal, regulated, client, HR, legal, medical, financial, or proprietary details that need redaction (optional but recommend declaring upfront)

procedure

1. confirm source and scope (input: pasted text; output: source type classification and completeness assessment)

   • identify whether input is a complete transcript, partial chat log, fragmented notes, or mixed source.
   • state explicitly what is present (full discussion, one-sided recap, summary, etc.) and what is missing or unclear.
   • do not guess at missing context; mark gaps as "context missing" or "incomplete record."

2. assess and redact sensitive content (input: transcript and sensitivity flags; output: redacted working version if needed)

   • if the user plans to send text to an external AI tool or the text contains sensitive details, create a redacted version first.
   • replace specific names, account numbers, addresses, health details, financial figures, legal case names, HR information, and proprietary data with placeholders like [PERSON_1], [CLIENT_A], [ACCOUNT_X], [PROJECT_Y], [DATE_1].
   • keep enough context to extract roles, relationships, and task logic (e.g., preserve "[PERSON_1] (VP Sales)" not just "[PERSON_1]").
   • if the user cannot safely share full text, ask for a summary they are allowed to provide instead.

3. extract candidate actions (input: transcript or redacted version; output: raw list of action candidates with verb-led statements)

   • scan for statements that imply a task, commitment, follow-up, request, blocker resolution, review, delivery, scheduling step, or handoff.
   • capture direct commitments ("I will send the report by Friday") and implied tasks ("we need to check with legal").
   • extract each action as a simple verb phrase (e.g., "send Q4 report to [PERSON_1]", "verify budget approval with [DEPT_X]", "schedule follow-up meeting").
   • note the line number or cue in the transcript where each action appears.

4. assign evidence and confidence levels (input: candidate actions and source text; output: action list with source cues and confidence tags)

   • for each candidate action, record the exact or paraphrased cue from the transcript that supports it.
   • label confidence as "confirmed" (explicitly stated commitment), "inferred" (strong signal but not explicit), or "needs verification" (ambiguous or conditional).
   • example: action "schedule demo with [CLIENT_A]", source cue "[PERSON_2]: we should set up a demo sometime next week", confidence "inferred" (no date or owner yet stated).

5. capture owners and dates carefully (input: action list and transcript; output: action list with owner and date fields populated where explicitly supported)

   • extract owner only if the transcript explicitly assigns the task (e.g., "[PERSON_1]: I will handle this" or "[PERSON_2]: can you take this?").
   • extract deadline or due date only if the text states a specific date, relative date ("end of week"), or clear target.
   • if owner or date is absent, ambiguous, or conditional, mark that field as "needs verification" instead of guessing.
   • do not infer ownership from seniority, speaking order, or assumed convenience unless the user explicitly requests a best-effort draft and you label every inferred item as such.

6. separate decisions from action items (input: action list and transcript; output: distinct decisions list separate from tasks)

   • identify confirmed decisions: statements that resolve a choice or direction (e.g., "we decided to move the launch to Q2", "we will use [VENDOR_A]").
   • do not convert discussion, brainstorming, suggestions, or debate into decisions unless the text explicitly confirms a decision was made (e.g., "we agreed to", "decision is", "let's go with").
   • list decisions separately with supporting quote or cue and any action items that depend on the decision.

7. identify open questions, risks, and dependencies (input: transcript; output: three separate lists)

   • open questions: explicit unresolved points (e.g., "[PERSON_1]: do we have budget for this?" left unanswered).
   • risks and blockers: stated concerns, missing prerequisites, or conditional statements that could prevent action completion (e.g., "blocked on legal review", "depends on vendor response").
   • dependencies: actions or decisions that one task depends on (e.g., "action X depends on decision Y being confirmed").

8. build the structured action register (input: confirmed actions with owners, dates, confidence, evidence, dependencies; output: formatted table or register)

   • create a table or list with columns: action, owner, due date, source cue, dependency, confidence, verification needed.
   • keep rows concise and scannable; expand in a separate section if needed.
   • sort by due date (earliest first) or by owner, as the user prefers.
   • include only actions with at least "inferred" confidence; flag "needs verification" items visibly.

9. add manual verification checklist (input: completed register; output: checklist for user review)

   • provide a checklist to prompt manual review before sharing or importing the register elsewhere.
   • checklist items: confirm each owner agrees to the task, verify each date is realistic and agreed, check that task wording matches what was actually discussed, confirm decisions are truly decided, verify redactions are complete and appropriate, confirm no sensitive content remains for external sharing.

decision points

• if the transcript is incomplete or fragmented: state explicitly what is missing and ask the user for clarification or additional context. do not fabricate attendance, decisions, or consensus.
• if sensitive content is present and external AI use is planned: create a redacted version and ask the user to approve the redactions before proceeding. if the user cannot share the text safely, work from a summary instead.
• if an action owner is ambiguous or unassigned: mark the action as "needs verification" rather than assigning it to the most senior person or the person who suggested it. offer to create a best-effort draft only if the user explicitly requests it and agrees to verify every inferred owner.
• if a deadline is missing or vague: mark the action as "needs verification" and list it in the "items needing manual verification" section. do not infer a deadline from surrounding context unless the user explicitly asks for a draft with inferred dates clearly marked.
• if the transcript contains legal, medical, HR, disciplinary, financial, or safety-critical content: recommend human review by qualified personnel before any action is assigned or communicated. do not proceed without the user's explicit acknowledgment of the sensitivity.
• if the user requests a reusable prompt for external AI: provide the template at the end of the output, but only after redaction is complete and the user approves the redacted version.

output contract

structured action register output includes these sections in order:

1. transcript source snapshot: one-line summary of source type, approximate length, date (if known), participant count (if known), and any gaps.
2. privacy and redaction notes: summary of what was redacted and why (or "no redaction needed").
3. action register table: columns are action, owner, due date, source cue, dependency, confidence, verification needed. one row per action.
4. confirmed decisions: bulleted or table list of confirmed directional choices with supporting cue and any dependent actions.
5. open questions: list of unresolved points from the transcript with context.
6. risks and dependencies: list of stated blockers, dependencies, and conditional factors.
7. items needing manual verification: list of actions, owners, dates, or decisions that require user confirmation or follow-up.
8. optional follow-up message or import notes: draft message or guidance for communicating the register to the team or importing into a task system (if user wants it).
9. verification checklist: bulleted checklist of steps to take before sharing or importing the register.

all output uses simple markdown formatting for readability. tables are plain markdown tables where supported; otherwise, use bullet lists or CSV-style text.

data format: plain text, markdown, or CSV (user preference). file location: none; output is pasted directly into the conversation or provided as copy-paste text ready for import.

outcome signal

the user will see:

• a visible, scannable action register with concrete verbs, clear owners (or "needs verification" flags), and stated deadlines or "needs verification" flags.
• separated sections for decisions, open questions, and risks that the transcript actually contains, not fabricated items.
• a clear "verification needed" marker on any action, owner, date, or decision that is ambiguous or inferred.
• source cues or transcript snippets that support each action, so the user can spot-check the extraction against the original text.
• a checklist reminder to verify owners, dates, and redactions before sharing the register with the team or importing it into their task system.
• confidence levels (confirmed, inferred, needs verification) on every item, so the user knows what to trust immediately and what to follow up on.

the register should feel like a draft ready for review, not a final assignment. the user takes the next step to confirm owners and dates before communicating or importing.

optional reusable prompt for external AI use

if the user wants a standalone prompt to send the redacted transcript to an external AI tool, provide this template:

"extract a reviewable action register from the redacted transcript below. use only the provided text. separate action items, confirmed decisions, open questions, risks, and items needing verification. for each action item, include: action (verb-led phrase), owner (or mark 'needs verification'), due date (or mark 'needs verification'), source cue (exact or paraphrased line from transcript), dependency (other actions or decisions it depends on, or 'none'), confidence (confirmed, inferred, or needs verification). do not invent owners, dates, decisions, or consensus. if information is missing or ambiguous, mark it as needs verification instead of guessing. format output as a markdown table or list for easy review."

safety boundaries

• work only from pasted or user-provided text. do not access private systems, files, recordings, calendars, email, messaging tools, meeting platforms, or external APIs.
• do not fabricate actions, owners, deadlines, decisions, quotes, attendance, or consensus. if the text does not support a claim, mark it as "needs verification."
• redact all confidential details before external AI use, including private names, client names, account numbers, addresses, health details, financial figures, legal case names, HR information, credentials, and proprietary data.
• do not ask for passwords, credentials, one-time codes, secrets, private keys, or unnecessary personal data.
• verify owners and dates manually with the team before assigning work or sending a recap.
• for legal, medical, HR, disciplinary, financial, security, or safety-critical transcripts, recommend qualified human review by appropriate personnel before any action is taken.
• preserve user privacy and confidentiality. if the user asks to use an external AI tool, ensure redaction is complete and approved before proceeding.

quality bar

a strong result gives the user a visible, reviewable action register that separates confirmed work from uncertain items, preserves source support for each action, protects sensitive details, and makes manual owner and date verification unavoidable before sharing or importing. the register should feel like a trustworthy first draft, not a final source of truth.

credits: original skill authored by Bell (design). enriched for implexa standards.