Coco Feng冯靖婉

31 people across 12 cohorts

Background

Leadership set a goal of company-wide AI adoption. An earlier rollout built on a single SaaS tool had stalled for nearly a year, and staff still used AI mainly as a chat window, with little effect on day-to-day productivity.

Challenges

Most staff had no working concept of an agent, and motivation varied; client data security was leadership’s first concern and a precondition for any proposal.

Solution

• Discovery: interviewed each team on current AI use and pain points, and traced the stalled rollout to that gap in understanding: people did not know what work an agent could take on.
• Rollout strategy: replaced the top-down push of a single tool with a small pilot that produced repeatable examples, followed by a phased rollout.
• Tool evaluation: benchmarked MaaS platforms with data security as the primary criterion, and wrote the evaluation report that leadership approved.
• Curriculum and teaching: designed a course around real business workflows and taught every session; revised it over several rounds of pilot feedback, breaking core concepts into progressive modules and opening with polished real-world examples.
• Usage tracking: set up LiteLLM with colleagues to issue model access centrally and track each person’s usage.

Impact

The course has reached 31 people across 12 cohorts. Agents are now part of colleagues’ daily work, and several have built their own tools; one copywriter’s tool, for example, gathers competitors’ past copy and adapts its tone for new drafts.

10–20 minutesmore than ten clients

Background

Account teams produced a monthly social media report for every client, each taking about 35 hours from data pull to revisions. A project to automate the reports had failed to ship for nearly a year.

Challenges

The company ran three databases, some built in-house and some licensed, with overlapping data but no connection between them; every client required a custom layout; and the users had no technical background.

Solution

• Data audit: took over the project, mapped the data assets account teams rely on, identified the items that needed regular refreshes, and updated them with the data team.
• Development: built the tool solo with Claude Code, from kickoff to working version in two weeks; a shared engine with per-client configuration handles every custom layout, and every output is quality-checked automatically.
• Launch and rollout: launched at the end of September, trained the full account team, coached them through adoption, and handed day-to-day use to the business.

Impact

A single instruction now produces a report, cutting each one from about 35 hours to 10–20 minutes. The tool covers more than ten clients in China and produces every current client’s monthly and quarterly reports; the three databases now sit behind one company-wide entry point.

Background

Generative AI is becoming a new way for people to learn about brands. DLG China began exploring a GEO service line in July 2026, aimed at making brands visible and credible in AI answers.

Solution

• Market and academic research: mapped the offerings and pricing logic of GEO vendors in China, reviewed and organised the research literature, and compared the sources Chinese AI platforms cite.
• Method design: designed a monitoring question bank and user personas organised by industry and decision stage, and set grading rules for the credibility of cited sources.
• Data collection: wrote collection scripts that export each industry’s answers and cited sources from an AI-answer monitoring platform as data ready for analysis.

Impact

Provided the technical groundwork for designing the GEO service line.

6 callstwo thirds

Background

Output from agents was not managed centrally: files went to management over WeChat with no production record; versions were scattered and could not be found months later; proposals for the same client did not always agree. In October 2026 the company decided that work would live in Google Drive, with one source of truth per client.

Challenges

Client folders had grown over the years in different layouts, and a full migration would cost more than it returned; the connector’s security rules barred the agent from uploading, moving, deleting or sharing files; contracts were visible to some colleagues only.

Solution

• One-command filing: designed /publish, which a colleague types when work is done; the agent works out the client, folder and standard file name, writes a summary and page index, and files on one confirmation, listing options instead of guessing when unsure.
• No conflicting files: set one naming and version rule; no file is overwritten silently, every copy is verified, existing files are never moved, renamed or deleted, and each client has one source of truth.
• Tiered disclosure: designated the account lead as owner of each client’s source of truth, with others proposing changes; when a contract cannot be found the agent notes it as possibly not shared rather than absent, and never infers its terms; each task uses one client’s material; the filing record stays internal.
• Connector and agent: built the solution with a colleague using the company’s own Google connector, writing only to company and client shared drives; the agent reads the source of truth and folder index before opening files; the skill syncs to colleagues’ computers through the connector each week.

Impact

Without the skill, an agent took 47 tool calls to find one client’s contract and quotation, opening other clients’ folders along the way; following the skill took 6 calls, all inside that client’s folder. In a read-only test on October’s monthly reports, two thirds of clients resolved to a filing folder with no question asked.

one or two nights

Background

Otome games run on large volumes of script, CG and music. Each proofreading and risk-review pass of a million-character script took three to five writers and me two to three weeks, reading chapter by chapter, and every CG needed an approved composition draft before it could be assigned.

Challenges

The company had no AI experience to draw on, and hand-drawn art is the core asset of an otome game, so AI could only support the process.

Solution

• Learning and building: learned agent workflows alongside project-management duties, then wrote the requirements and PRDs and built the tools.
• Scripts: built an automated proofreading pipeline that flags risks reviewers tend to miss and lists them by scene, leaving every edit to the writers.
• CG and music: batch-generated composition references from character sheets and costume art for artists to redraw, so final art stayed hand-drawn; for music, evaluated open-source models, then switched to programmatic MIDI references.
• Securing resources: wrote the resource request with per-image cost estimates and usage controls, and obtained model accounts and access to project files.

Impact

Script review went from two to three weeks to one or two nights; CG pre-production moved from reference hunting and hand sketches to batch-generated composition references; the music references set the tonal direction for pre-production. On the strength of this work, I moved from project management to AI engineering lead.