Felo Skills for Doubao Work: Generation Is Solved, Verification Is the Job
Add Felo skills to Doubao Work for sourced research and slide drafts your scheduled tasks can run — with an evidence step before the file ships.
Doubao Work (豆包工作) does not answer a question. It finishes a job. Describe the outcome and it plans the task, draws on a library of domain 技能 (skills) covering investment, medical, creative, and e-commerce work, and returns a real deliverable: a deck with layout, a spreadsheet with formulas, a document with structure, a website, an image.
It also closes loops most assistants leave open. 定时任务 (scheduled tasks) run the same job on a cadence without anyone opening an app. Computer use lets the agent read and edit local files, open pages, and drive a local CLI when that is the shortest path to a result. And with documents and knowledge already living inside your workspace tools, the distance between research and delivery is short.
What the pipeline cannot tell you is whether the numbers are true.
Generation is solved. Verification is the bottleneck. That is not a complaint about the platform — it is the shape of any workflow that ends in a file. A generated report that cites nothing is a liability the moment it reaches a client or a manager. A market number recalled from training data is worse than no number at all, because it looks authoritative.
Felo skills are the evidence step. They search with dated sources, fetch the actual pages, run a longer investigation when a question deserves one, and hand findings back in time for the writing stage. The deck still comes from Doubao Work. The claims now trace to something.
Three seams where that matters: before drafting, so the outline starts from facts instead of memory; inside a scheduled task, so the research refreshes and diffs against last week; and before delivery, so source notes travel with the file.
Two ways to connect Felo to Doubao Work
Pick the path your build supports. Both end in the same place: a research step the agent can call during a task.
| Path A — an imported bundle | Path B — the Felo CLI | |
|---|---|---|
| Setup | Import a skill folder through the 技能 / 连接器 (skills / connectors) surface | npm install -g felo-ai once, then store the key |
| Availability | Depends on your build — personal, team, or enterprise | Any build whose agent can use your computer |
| Who runs it | The agent, during a task, on your instruction | You at a terminal, or the agent through computer use |
Path A — import a skill bundle, where your build exposes one
Skill bundles are folders with a SKILL.md inside, the same format Felo publishes in a public repository. If your version of Doubao Work shows a skill or connector import surface, that is the shortest route: add the bundle, and the agent can select it the way it selects a built-in skill.
Skill and connector availability differs between personal, team, and enterprise builds. If the import entry is not visible on your account, use Path B.
Path B — run the Felo CLI as the research tool
Install the CLI once. Then either call it from a terminal during setup, or let Doubao Work's computer-use ability invoke it as part of a task on your machine.
npm install -g felo-ai
felo config set FELO_API_KEY your-api-key-here
felo search "latest industry policy"
felo slides "Q3 competitor review, 8 slides"
The agent installs nothing and learns no new commands. It runs a command that already exists, and the result comes back into the task. The key stays in your user config rather than in a task prompt.
Tell the task what a good answer includes
Whichever path you take, the prompt carries the standard. Source requirements typed into the request are what keep the output reviewable.
研究这个行业最近三个季度的政策变化，整理成一份报告和一套幻灯片。 每条结论必须给出来源和日期；无法核实的部分标注「待确认」；结尾附来源清单。
In English: research this industry's policy changes from the last three quarters, produce a report and a deck, cite a source and a date for every conclusion, mark anything unverified as "to confirm," and finish with the source list.
Three instructions are doing the work there. The citation rule makes claims checkable. The 待确认 (to confirm) label keeps a gap from being filled with something plausible. The closing source list turns the file into something a reviewer can audit instead of reading on trust.
Five Felo skills worth importing
Two read the outside world, two produce the deliverable, one supplies the images.
Felo Search — live web answers with dates
The research step everything else leans on, and the same engine that powers felo.ai/search, now reachable from inside a task. Findings arrive dated and linked, which is also what makes a scheduled run diffable.
Search this quarter's policy updates for the industry, list them by publication date, and give the source for each.
Felo Web Fetch — pages into structured data
Search tells you a page exists; fetching tells you what it says. Pricing pages, policy documents, and product listings come back as reusable structure rather than a paragraph someone has to re-read.
Fetch these five competitor pricing pages and return the tiers and prices as a table.
Felo Content to Slides — notes and URLs into an outline
The outline stage, as text. This is the version worth arguing about: changing six headings costs nothing, and regenerating a finished deck costs a review cycle you did not budget for.
Turn the findings in this task into a deck outline I can edit before anything is generated.
Felo Slides — the deck itself
Once the outline holds, this builds the file. For a scheduled run the point is the ending: the task stops returning chat text and starts returning something you can forward. The same deck builder runs standalone at felo.ai/slides when you would rather iterate outside the task.
Build an eight-slide deck from these findings and put the data sources in the speaker notes.
Felo SuperAgent — images in the same run
Plenty of Doubao Work tasks end in a picture — a product shot, a campaign visual, a logo. SuperAgent generates and revises those without leaving the task.
Create a product hero image with a clean background and the packaging as the focus, plus two alternates in other styles.
Beyond these five, the catalog at felo.ai/skills/doubao-work also carries Felo LiveDoc for answers grounded in your own documents, plus bundles for screenshots, social writing, and landing pages — the same set Felo publishes for every platform it supports.
Four tasks people already run this way
Each is written as the request, followed by what the Felo layer adds.
Scheduled competitor watch
Every Monday morning, someone needs a summary of what competitors changed: pricing, packaging, launch posts, new case studies.
Every Monday at 8:00, check these competitors' pricing pages and public updates, then list what changed and what to watch this week.
A recurring brief that flags what changed since the last run instead of re-describing the same pages, with a link for every change. The diff is what keeps a weekly brief worth opening in week ten.
E-commerce launch material
A new product needs listing copy, a comparison table, and a short campaign brief, all grounded in what the category currently sells.
Research how the top products in this category are positioned and priced, then draft our listing copy and a launch plan.
Copy and a plan built on observed category data, with the source list kept for the next iteration.
Medical and academic literature pass
A clinician or researcher needs recent studies on a treatment question, sorted by evidence quality.
Find the multicentre studies on this treatment from the last five years and summarise efficacy, safety, and population, with the paper for each point.
A literature summary where each line traces to a paper, ready to be reformatted into Doubao Work's document output. This is the case where a plausible-sounding citation is worse than a missing one.
Briefing pack from a long report
An internal report exists; the meeting needs ten slides and a one-page summary.
Turn this report into a ten-slide deck and a one-page summary, keeping the key numbers and their sources.
A slide draft and a summary generated from the report, with the source notes preserved — so when someone asks where a figure came from, the answer is already in the file.
Schedule it only after it survives one manual run
The order matters more than the tooling.
Run the research once by hand. Ask the question, read the sources, decide whether the answer holds. A timed task that produces unchecked claims scales the mistake, not the value.
Decide the deliverable before you automate. Deck, document, or sheet. A recurring job that ends in a different shape every week cannot be compared against its own history.
Add the image and video skills after the text path is stable. Plenty of tasks end in something visual or in a recording nobody wants to rewatch. Those are the easy second step.
Then move it to a scheduled task. Once the manual run is stable, scheduling is nearly a one-line change — and every run should diff against the previous one, with a person still on the numbers and quotes.
Schedule this research process for every Monday at 8:00 and diff the result against last week.
Guardrails
Ask for source lists by default. A report without sources cannot be reviewed. Put the requirement in the recurring task definition so every run inherits it.
Separate facts from recommendations. Generated analysis blends what was retrieved with what the model advises. Ask for a summary that labels which sentences come from sources and which are inference — the second category is where a confident sentence most often needs a human.
Keep the API key on your machine. The Felo CLI stores credentials in your user config. Do not paste a key into task text, where it becomes part of the conversation record and travels with anything you share.
Re-check anything that triggers spending or a public post. Computer use and scheduled tasks can act while nobody is watching. Money, publishing, and messaging steps belong behind a manual approval — as a standing rule, not a one-time setting.
FAQ
Does Doubao Work support importing third-party skills?
It depends on your plan — personal, team, or enterprise — and on what your account has enabled. If the skill or connector import panel is visible, you can import bundles that follow the SKILL.md format. If it is not, the Felo CLI installed on your machine does the same research job as an external tool.
Can a scheduled task call external research?
Yes, as long as the task returns a stable result when you run it by hand. Run it once, check the sources, then schedule it and make every run diff against the previous one.
How are Felo skills different from the ones built into Doubao Work?
The built-in skills cover a wide range of professional tasks and usually return the document, spreadsheet, or deck itself. Felo skills handle external retrieval, page extraction, and sourced research. They are complementary: one produces the deliverable, the other produces the evidence.
Does this cost extra?
Felo search and generation use your Felo API quota, while Doubao Work bills its own models and compute under its own rules. A deep research run spends more quota than a single search, so confirm the value with a manual task before you automate it.
Can the generated content be published as it is?
Treat it as a first draft. Sourced output lowers the cost of checking, but a person still has to confirm the conclusions, the numbers, and the citations before publishing — especially for pricing, medical, and compliance material.
Which skill should I start with?
Start with Felo Search when stale information is the problem. Add SuperAgent when a task ends in an image. Once both are in place, add a slide skill for whatever needs a deck.
Add one research skill, run a task you already do, and compare the output against the last version you shipped. If the difference is that a reviewer can spot-check the file without asking you a question, the loop is closed — and the job is ready for 定时任务.
Explore Felo skills for Doubao Work →
This post is also available in 简体中文, 日本語, 한국어, 繁體中文, हिन्दी, Français, العربية, Русский, اردو, Bahasa Indonesia, Deutsch, Tiếng Việt, Türkçe, Italiano, ไทย, Español, বাংলা and Português.