agentic-workflows-portfolio

Mesas de Cooperación. Conservación Internacional

My consulting work for Conservation International Foundation Ecuador, reinvented: instead of chatting with AI, I now work inside the project files with Claude Code: producing better deliverables, dashboards for the Cooperation Tables in Sucumbíos and Orellana, and email automation for the two-year roadmap.

hero

The Problem

I’m the consultant behind the technical accompaniment to establish two Mesas de Cooperantes (Cooperation Tables) for conservation and sustainable development in the Amazonian provinces of Sucumbíos and Orellana. Contract runs October 2025 – April 2026. The work produces long, dense deliverables in Spanish (.docx reports, roadmaps, governance models, constitutional acts), plus the ongoing coordination of many stakeholders: provincial governments, NGOs, technical secretariats, and management groups. The provincial governments led the tables; my part was the method, the facilitation and the instruments.

Before Claude Code, I was using AI through chat interfaces: great for brainstorming, bad for real project work. Context was always lost. Every conversation started from zero. Feedback from the client had to be manually translated into edits. Deliverables lived in one place, my AI assistant lived in another, and I was the glue.

The Solution

I moved the consultancy itself into Claude Code. The project folder holds every contract, every prior deliverable, every feedback document, every roadmap. Claude reads them directly: no re-explaining context, no copy-pasting.

What this unlocks:

Deliverables get written with full context. The main report (Entregable 5: the Final Consolidated Report) is edited inside its own .docx file, preserving formatting, comments, and the table-of-contents structure that Word demands.
Client feedback becomes a plan. When the client sent back 14 review comments, Claude Code parsed them from word/comments.xml, grouped them into 5 phases, and worked through each one, with me approving every change.
Roadmaps become dashboards. The two-year roadmaps for each province (Hojas de Ruta 2026–2028) are being turned into interactive HTML interfaces where stakeholders can see their assigned actions, mark them done, and stay aligned.
Follow-ups become automated. Emails to the Grupo Gestor, reminders to the Secretaría Técnica, status requests to the Procurador Síndico, all templated and driven by the same structured data behind the dashboards.

The bigger shift is methodological: my consultancy is now a repository, not a series of chat threads.

What Changed Since April 2026

The monitoring system is built and ready, and not yet deployed with the members. It was prepared for the launch workshop of the Sucumbíos provincial government on 25 April 2026. The public views are read-only and open to anyone: a landing page with live stats across both provinces, plus one panel each for Orellana and Sucumbíos with a scorecard, a filterable activity table and a timeline by sub-table.

Each member would get a personal link. A link with a unique token opens a “Mi trabajo” (my work) tab with the member’s agenda, streak, quarterly target and their sub-table’s standing. There is no way to pick an identity from the page, so nobody can act as someone else.

Reports are verified in one click. Members submit progress through a form. The technical secretariat verifies or rejects each report from its inbox or from Mi trabajo, and rejected reports link straight back to the activity. Reminder and notification emails go out from Google Apps Script.

Built to be trusted with public data. A live “Informe” (report) tab in both dashboards, CSV export, quarterly report pages, a data-sync command, and tests for priority scoring, the member states and the keyword matching that assigns activities to sub-tables. A May pass fixed data errors, Gmail compatibility and failures that had been silent.

Screenshots

Interactive dashboard for the 2026–2028 roadmap of one of the provinces: each action has an owner, a due date, and a status, and members of the Mesa can log in to track their part.

Client feedback parsed out of word/comments.xml and turned into a 14-action, 5-phase implementation plan.

Result

Entregable 5 (Final Consolidated Report, 7 sections + annexes) produced inside Claude Code with full formatting preserved.
Client feedback loop tightened: 14 review comments processed structurally instead of hunted for in a Word doc.
Dashboards for the Cooperation Tables’ two-year roadmaps, in progress, replacing the static spreadsheet view.
Email automation scaffolded for stakeholder communications (in progress).
Methodological shift: from chat-based AI assistance to project-based, context-rich work. This is reusable across every future consultancy engagement I take on.

Key Decisions

Structured docx editing over export/reimport. Instead of exporting to markdown, editing, and reimporting (which destroys formatting), edits happen directly on individual runs within paragraphs using python-docx + lxml. Tables of contents, comments, embedded images, and track changes all survive.
Comments parsed as first-class input. Word comments live in word/comments.xml and are invisible to python-docx. I wrote a small parser that turns them into an actionable list. That became the bridge between client feedback and implementation.
Dashboards as the “living” deliverable. A .docx is a snapshot; a dashboard keeps working after the contract ends. The dashboards are the handoff that gives the Mesas something to use, not just read.
Claude Code as a co-pilot, not an author. Every major edit is reviewed before it touches the deliverable. The client gets consulting work, not AI output.

Under the Hood

Stack: Claude Code, Python 3.13 (python-docx, lxml, zipfile), HTML/JS for dashboards, macOS textutil for quick text extraction.

The project folder functions like a small knowledge base: contract documents, previous deliverables (E2–E4), feedback files, and working drafts all sit side-by-side. Claude Code is instructed (via CLAUDE.md) on which files are targets for edits, which are reference-only, and the technical gotchas of editing .docx XML without breaking Word’s rendering.

How to Demo It

Open the project folder in Claude Code: show CLAUDE.md, the contract, the prior deliverables, and the main report side-by-side.
Show a real client comment in Entregable 5.docx and the feedback-implementation plan that came out of it.
Open the roadmap dashboard for one of the provinces: click through a stakeholder view.
Explain the shift: “This used to be five ChatGPT tabs and a lot of copy-paste. Now it’s one project, one context, one co-pilot.”

Limitations & Setup

Everything is in Spanish: deliberate, the client is Ecuadorian.
.docx editing is fragile: always save-close-reopen between major edit phases, never modify paragraphs containing <w:drawing> or <w:pict>.
After heading changes, the Table of Contents must be manually refreshed in Word (F9 on Windows, manual right-click on Mac).
The monitoring system is built and not yet deployed with the Mesa members. Sucumbíos was wired first; Orellana runs on the same code with its own data.
Client data is confidential. The project folder itself stays private; only this portfolio doc is public.

Demo Pitch

“I’m a consultant working for Conservation International in the Ecuadorian Amazon. Claude Code replaced my chat-based AI workflow with something that lives inside the project itself: every contract, every deliverable, every comment from the client. It doesn’t just help me write; it helps me run the whole engagement. The dashboards and email automation are how I’m turning a report into something the stakeholders can actually use.”

Changelog

2026-09-28

Added “What Changed Since April 2026”: monitoring system built and ready (not yet deployed), prepared for the Sucumbíos workshop (25 April 2026), personal token links, one-click verification, Apps Script emails, live report tab, tests.
Updated stack, effort, and the limitation on rollout.

2026-04-16

Initial portfolio doc