Meeting → Tasks

Coordenadas

SOP — Meeting Transcript
to Tasks

Turns a raw meeting transcript into a cleaned meeting record and a deduplicated, Coordenadas-scoped task backlog — with a project manager's judgment about what is and isn't our work.

OwnerPablo (Coordenadas)

TypeSemi-automated (skill)

Trigger/meeting-to-tasks or auto

Outputsmeeting record + TASKS.md

Skillmeeting-to-tasks

Run it

How to run it

This is all you need to use it. The rest of the detail is below.

Slash command

/meeting-to-tasks <ruta o pegar la transcripción> [marca] [fecha]

With empty arguments, the skill asks for the transcript and the brand before doing anything.

Auto trigger and location

Automatic: just hand over a transcript or meeting notes.
Skill: ~/.config/opencode/skills/meeting-to-tasks/.
Canonical: docs/SOP_reference/SOP_Meeting_Transcript_to_Tasks.md.

Overview

At a glance

Every client meeting produces two durable artifacts:

A cleaned meeting record — Summary, Notes, Transcript and Action Items, saved under the brand's knowledgebase.
Tasks merged into the brand's TASKS.md — only work inside what Coordenadas offers the client, deduplicated against what already exists.

The skill reads the brand's real context before deciding what is a task, and it never invents facts the transcript and context do not support. The result is a traceable memory of every meeting, not a note that gets lost.

Inputs

What you need before starting

The minimum for the skill to run end to end.

Transcript — pasted in chat or as a file path.
Extra notes (optional) — your own meeting notes; merged into the Notes section.
Brand — the Brands/<Brand>/ folder.
Meeting date — from metadata; if absent, the skill asks.
Participants — the transcript's speakers; if only some captured, the full list is requested.

One-sided transcripts

If participants were muted during a screen-share, the transcript captures only one person. The skill then leans on your notes and flags in the Summary what could not be reconstructed — it does not fill the gaps on its own.

Process

The workflow, step by step

Eight steps, from the raw transcript to the updated backlog.

Save the raw transcript

Copy the source verbatim into meetings/raw/ before any cleaning and verify the copy.

Load the brand context

In order: AGENT.md → CONTEXT.md → marketing brain → knowledgebase 00–07 → rules/ → sitemap → current cycle → root AGENTS.md.

Confirm the two underivable facts

Meeting date and full participant list. If they can't be derived, they're asked — never assumed silently.

Write the cleaned record

Sections in fixed order: Summary → Notes → Transcript → Action Items. In the transcript's language. Clean, don't rewrite.

Extract the action items

Each item with action, owner, side, priority and target date.

Apply the scope filter

Coordenadas work becomes a task; client-side and out-of-scope items go to Fuera de alcance.

Merge into TASKS.md

Dedupe by action + owner, assign the next ID, update statuses, log the change.

Report

Files written, IDs added/updated, out-of-scope items, and any assumption made.

Outputs

Files and naming

Everything lives under the brand's knowledgebase, with TASKS.md at the root.

Meeting record

Naming convention

<Brand>_Meeting_YYYY-MM-DD_<slug>.md

Factronics_VE_Meeting_2026-09-11_ghl-setup.md PayWise_Meeting_2026-09-11_q3-content-review.md

The date is the meeting date, not the processing date; ISO sorts chronologically. The raw copy lives beside it, in raw/.

Folder structure

Brands/<Brand>/
├── TASKS.md  ← backlog vivo, fusionado
└── 01_knowledge_base/research-and-notes/meetings/
    ├── <Brand>_Meeting_YYYY-MM-DD_<slug>.md   ← registro limpio
    └── raw/
        └── <Brand>_Meeting_YYYY-MM-DD_<slug>_raw.<ext>

TASKS.md schema

IDs <CODE>-NNN, monotonic and never reused.

Statuses: Pendiente · En progreso · Bloqueada · Hecha · Descartada · Fuera de alcance.

TASKS.md is a single living file at the brand root, always in the brand's TARGET_LANG, even though the meeting record follows the transcript's language.

Scope

What counts as a Coordenadas task

Source of truth: root AGENTS.md (Agents 1–10 + W1–W2) plus the brand's AGENT.md/CONTEXT.md.

In scope

Monthly content cycle (plan, carousels, blogs, infographics, newsletters)
Image production and prompts
Landing pages and marketing hub
GoHighLevel: subaccount, bot, pipelines, workflows, automations
Performance reports and scheduling
Brand, strategy, assets and rules

Out of scope

The client's own operations (uploading their contacts, editing their site)
Platform/vendor matters handled by the client's own staff
Speculative ideas with no owner or decision
Work blocked by a client decision

Only Coordenadas work gets a task ID. Client work is listed as a dependency under Fuera de alcance, so it isn't lost without polluting the active backlog.

Principles

Principles to preserve

Never invent. If it isn't in the transcript, the notes, or the brand context, it isn't a fact.
Preserve verbatim names, numbers, prices, dates and product models.
Brand hard rules bind the output — e.g. Factronics VE: zero Linx.
Language split: record follows the transcript; TASKS.md follows the brand's TARGET_LANG.
No false completion: only marked Hecha with evidence.
The raw copy is always kept. The cleaned file never replaces the original.
Traceable trail: every task links back to the meeting it came from.

Mistakes

Common mistakes

Mistake	Fix
Using the processing date as the meeting date	Derive it from metadata; if absent, ask.
Writing tasks for client-side work	Client items go to Fuera de alcance.
Creating a task that already exists in TASKS.md	Dedupe by action + owner; update the existing row.
Rewriting the transcript in your own words	Clean only: label speakers, drop filler, keep the wording.
Skipping rules/	Brand hard rules bind the task wording.
Replacing the original with the cleaned file	Always keep both; the raw lives in raw/.