---
title: Build a personal research agent that turns saved tutorials into topic cards
description: >-
  Follow a real editorial task from six sources and an article inventory to
  actionable topic cards. Includes actual outputs, reusable requests and a
  bilingual working pack.
canonical_url: 'https://onevium.com/blog/personal-agent-research-topic-cards'
language: en
published_at: '2026-11-08'
publisher: Onevium
---
# Build a personal research agent that turns saved tutorials into topic cards

Follow a real editorial task from six sources and an article inventory to actionable topic cards. Includes actual outputs, reusable requests and a bilingual working pack.

## Turn a saved tutorial into a next action

A bookmark folder full of AI-agent tutorials creates a familiar problem: which one should you try, what useful work will it do, and have you already covered the same idea?

While preparing practical Onevium articles, we used a file-based research assistant to turn source notes, a reader brief, and an existing article inventory into actionable topic cards. We kept two directions and deferred two others with specific reasons. **The deliverable was a decision about the next piece of work, with evidence for that decision.**

This article records work completed on November 8, 2026, in Codex. Two Claude Code attempts produced no usable artifact, so this is not a successful Claude Code trial or a comparison of the products. The materials, requests, and editorial outputs below come from the work that actually completed.

## One method worth borrowing from the tutorial

In Teacher’s Tech’s [Build Your First AI Agent](https://www.youtube.com/watch?v=gHB4JFG9i3k), the segment around [12:45](https://www.youtube.com/watch?v=gHB4JFG9i3k\&t=765s) plans a research procedure and saves it as instructions. Around [21:14](https://www.youtube.com/watch?v=gHB4JFG9i3k\&t=1274s), the presenter revises the existing report’s summary and comparison table.

We read the full automatic captions and borrowed the idea of saving a working method and revising an existing artifact. We could not verify the video footage and did not reproduce its research report. Specific terminology in automatic captions still needs checking against official documentation.

Our own assignment was smaller: choose two practical articles from the available material. Establishing whose decision the assistant supports makes the first workflow much easier to define.

## Prepare the material the assistant cannot infer

The source pack contains six records: four video references and two official Claude Code documentation pages. Each has an ID and a reading status. S1 has a fully read automatic transcript; S3 has only a public description and chapter list. Those are different evidence levels.

We also exported an inventory of **62 registered website articles**, including titles, descriptions, and stable slugs. That is the inventory size at the start of this task, not a search-engine indexing count. It helps locate possible overlap but cannot establish full-text duplication.

The brief specifies readers who want a personal workplace agent and asks for two small topics with concrete file-based outputs.

[Download the bilingual working pack](https://onevium.com/downloads/workplace-editorial-lab.zip). Start with its README, inspect the source notes and inventory under inputs, and compare the actual artifacts under outputs. The pack does not redistribute the video transcript or raw runtime logs.

```text
workplace-editorial/
  WORKFLOW.md                   Reusable working method
  inputs/source-notes.md         Sources and reading status
  inputs/published-articles.json Article inventory snapshot
  inputs/brief.md                Current assignment
  prompts/                      English and Chinese requests
  outputs/                      Topic cards, draft and review
```

The first exercise needs access to this folder. In any other environment, confirm that the assistant can read the files. A pasted URL does not establish that a page has been read.

## Specify the handoff you want

These are the constraints used to organize this assignment. Replace the audience and input material for your own work:

```text
Read WORKFLOW.md, the source notes, brief and article inventory.
Choose two practical topics for readers building personal work agents.
For each topic, return:
- the task the reader will complete;
- source IDs and what was actually read;
- the closest existing article and the new contribution;
- the concrete artifact to build;
- missing evidence and the next action.
Keep deferred directions and their reasons.
Do not label a topic tested or published.
The inventory has titles and descriptions only; investigate possible overlap.
Save topic-cards.md without modifying the inputs.
```

A saved card carries the decision into the writing assignment. The workflow file preserves the recurring method rather than the conclusion of this particular run.

To adapt the material to Claude Code, consult its [project instruction documentation](https://code.claude.com/docs/en/memory). Our pack uses an explicitly requested WORKFLOW\.md. An ordinary Markdown file does not execute itself and is not automatically an installed Skill.

## The four decisions we recorded

This is an edited presentation of the actual topic cards. The full Chinese artifact and an editorial English translation are in the pack.

| Direction                    | Decision | Reason and next action                                                            |
| ---------------------------- | -------- | --------------------------------------------------------------------------------- |
| Topic research assistant     | Proceed  | Read captions, source notes and inventory are available; produce actionable cards |
| Learning notes to an article | Proceed  | Use this task’s records; retain a first draft, review and revision                |
| Proactive scheduled workflow | Defer    | Related articles already exist; obtain a new mechanism and an actual task         |
| Mobile task assistant        | Defer    | Verify product entry points and local/cloud execution conditions first            |

Proceed means invest in the next step. At selection time, the second direction still had no first draft; that gap remained on its card until the later writing work.

The scheduled-workflow idea could still become useful. The site already has a guide to [automation delivery and recovery](https://onevium.com/blog/reliable-ai-automations-delivery-recovery), so another article needs a new reader problem and supporting material.

![Four actual editorial decisions: proceed with topic research and notes-to-draft workflows, defer scheduling and mobile tasks, and retain a next action for each.](https://onevium.com/assets/blog/workplace-editorial/research-en.svg)

A visual excerpt of this task’s editorial output, not a product screenshot. The complete artifact is in the working pack.

## Review whether each decision leads to useful work

We checked three things beyond the presence of source links.

First, evidence level. “20 minutes” in S3 is a video-title claim, not our setup measurement. S4’s description of phone access does not prove that a local task runs with the computer off.

Second, the contribution beyond existing content. Renaming a marketing-plugin overview as a marketing-agent article adds little. Our new deliverables are a working directory, actual selection decisions, and artifacts that continue into editing.

Third, an actionable next step. “Research agents further” is broad. “Use the notes and card to write an opening, then record the edits it needs” can start immediately. That became the [companion draft-and-revision exercise](https://onevium.com/blog/personal-agent-notes-to-reviewable-article).

We did not perform a blind evaluation or measure time saved. The verifiable result is the material, decision cards, and subsequent writing. A different team should try its own assignment before drawing wider conclusions.

## Reuse the method with new material

Replace the brief and source notes, update the inventory, and explicitly ask the assistant to reread them. Preserve “description only” when captions are unavailable. Where an existing article covers the topic, ask for a concrete new contribution or an update to that article.

Keep a decision and its evidence from every run. After several assignments, distinguish stable procedures from one-off conditions before packaging them as a Skill. Our [guide to maintaining team Skills](https://onevium.com/blog/maintain-team-skills) covers that next stage.

Start with three to five tutorials you have saved and one task you actually need to finish this week. Ask for a card that makes the next action clear.

Source: [Teacher’s Tech — Build Your First AI Agent](https://www.youtube.com/watch?v=gHB4JFG9i3k)
