DigiNomads
Ask the help desk
Sandbox build. This deployment is noindex and takes no payment; the public site is the production build.
Skill pack

Research desk

Four skills for source-checked research: a source ledger, a claim checker, a comparison matrix and a freshness scan.

Starter 4 skills Reviewed 10 November 2026 Markdown mirror
Skill

source-ledger

Trigger
When a task needs facts from more than one page
Output
A table of claim, source, access date, quote and confidence
Guardrail
A source the agent did not open is marked 'not opened', never cited.
Skill

claim-checker

Trigger
Before a draft with numbers, dates, rules or quotes is published
Output
Each claim labelled fact, opinion, experience or unknown, with the supporting sentence
Guardrail
A claim without a supporting sentence is removed, not softened.
Skill

comparison-builder

Trigger
When two or more options must be compared on the same criteria
Output
One row per option, one column per criterion, with 'not published' where data is missing
Guardrail
No single score or ranking; the trade-off stays visible.
Skill

freshness-scan

Trigger
When a page has a reviewed date older than the source's last update
Output
A list of claims whose source changed, with the old and new wording
Guardrail
A changed source is a changed claim; the page date moves only after a person re-reads it.
Install

Put the pack in the project

  1. Create one folder per skill under `skills/` in the project the agent works in: `source-ledger/`, `claim-checker/`, `comparison-builder/`, `freshness-scan/`.

  2. Add a `SKILL.md` to each folder with a specific `description`. The description is the trigger; write it as a situation, not a topic.

  3. Put long source notes in `references/` and keep the main file short enough to read in one pass.

  4. Run the three tests below with a prompt that should trigger each skill. If the agent never loads it, the description is too vague.

Folder layout 
skills/my-pack/
|-- SKILL.md
|-- references/
|-- scripts/
`-- assets/
Test

Prove the pack does something

Compare two visa routes for a freelance designer and cite the official pages.

Pass condition: The agent creates a two-row table with an access date and marks missing cells 'not published'; it does not rank the routes.

Here is my article draft. Check every number against these links.

Pass condition: Each number is labelled fact, opinion or unknown; at least one unsupported claim is removed rather than reworded.

This page was last reviewed in March. Is it still right?

Pass condition: The agent lists the sources that changed and does not update the reviewed date on its own.

Why this pack first

Research is where an agent's confidence is most expensive. A wrong date in a client update is embarrassing; a wrong eligibility rule in a move plan is costly. This pack makes the evidence visible before the prose gets smooth.

The four skills cover one loop: collect a source, check a claim, compare options, and notice when the source changed. They are deliberately small. A research skill that tries to do all four in one file becomes a prompt with a folder around it.

The rule behind the pack

Every factual sentence has a source a reader can open, or it is removed. That rule is easy to state and hard to follow, because a model can produce a plausible sentence faster than a person can check it. The skills put the check in the path instead of relying on willpower.

Adapting it

  • If your work is legal or regulatory, add a jurisdiction field to the source ledger and make it required.
  • If you research in more than one language, add the original-language quote next to the translation.
  • If you publish on a schedule, make the freshness scan a weekly task and keep the last run in the repository.

Failure modes

  • The ledger is longer than the article. That is fine for a first pass; trim the ledger to the claims that carry the argument.
  • Every claim is marked “opinion”. That is a way to avoid checking. Require a source for anything a reader could verify.
  • The model cites a search result. Ask for the page it opened and the sentence it used. A snippet is not a source.

Related

Sources

What this pack is based on

Agent Skills specification
Official documentation
Open
Anthropic public skills repository
Official documentation
Open
Claude Agent Skills overview
Official documentation
Open
Skills are instructions, not a security boundary. Read a skill before you install it, keep secrets out of the folder, and treat scripts as code you are about to run. The full pack is in the Markdown mirror at /skills/research-desk.md.

Found a gap in the library?

Send the scenario, the prompt that failed, or the harness change that broke your setup. The help desk takes questions and corrections from humans and agents, and every accepted fix updates the page date.

Ask the help desk Browse the prompt library