Operator guide
Nine review skills, in the open.
newsroom-skills is a public, MIT-licensed repository of the editorial-review skills T3 runs for financial brands, rewritten as Agent Skills any coding agent can read and follow. Point an agent at a draft and the skills check every source, trace every claim, screen for regulated language, and write an audit trail. A person reads the result and decides what ships.
This is v1 of the guide: what the repository is, how to run it, and what it does not do. The full guide is coming.
Who it is for
Content and marketing leads at banks, credit unions, brokerages, and fintechs, where a wrong number or a missing disclosure is a real problem.
Teams starting their work with agents who want a review layer they can read line by line before they trust it.
Teams already running one who want to hold their process up against a worked example and see where it is thin.
A starting point, not a system. The method is here; the knowledge that makes it work for your brand is yours to build.
What is inside
Nine skills in four tiers. The tiers are reading order, not architecture.
Front gate
intake-router
reads the request before any work starts and asks when the answer is ambiguous instead of guessing.
Sourcing and verification
source-registry
sourcing tiers, corroboration rules, lookback windows, and citation format. Ships the schema for a source list, not a populated one.
source-verification
every cited URL is fetched, live, on-topic, and inside the recency window, or it does not ship.
fact-check
the claim matches the source. Catches distorted framing, invented terminology, and stats that drifted from the data.
Editorial checks
compliance-gate
pattern checks for the language rules that govern financial advertising. A starting point for your compliance team to extend, not regulatory guidance.
banned-terms-check
runs a ban list against a draft. Small, forkable, and the first skill to try.
voice-check
the conversational-editor pass, read against a voice profile you supply.
editorial-standards
the shape layer. Required sections, section order, headline conventions, unit-of-writing limits.
Audit and handoff
run-log
what every editorial run records before it counts as done.
Four reference files sit at the root: the hard rules, the starter ban list, the altitude model (where a rule lives and who owns it), and what the repository does not do.
Run it
Clone the repository and open it in your coding agent. Claude Code and Codex register the skills on their own. The prompt names every skill by path, so it also works in an agent that has registered nothing. Paste this, with the path to your draft:
Review the draft at <path-to-draft.md> using the skills in this repository. Read style-guide.md and banned-terms.md first. Then read and apply each skill in this order, one at a time: skills/intake-router/SKILL.md, skills/source-registry/SKILL.md, skills/source-verification/SKILL.md, skills/fact-check/SKILL.md, skills/compliance-gate/SKILL.md, skills/banned-terms-check/SKILL.md, skills/voice-check/SKILL.md, skills/editorial-standards/SKILL.md, skills/run-log/SKILL.md. Where a skill needs a tenant file this repository does not ship, do what the skill says to do in that case, name the gap in the run log, and continue. Do not invent the file. Fetch every URL you check; a link you have not fetched is unverified. Finish with the marked-up draft and the run log the run-log skill specifies. Flag, do not silently rewrite. Every change is a person's decision.
No draft yet? Point it at `examples/compliance-review/draft.md` and compare your output with the after state and run log beside it. A second prompt in `AGENTS.md` reads the tree in order and reports what your own process is missing.
What it does not do
From `not-in-v1.md`, unedited:
It does not compose the skills into a workflow. There is no router that decides which skills run for which request, and no loop that runs them in sequence with retries and gates. You run them in order, one at a time, or you build the composition yourself.
It does not ship a populated tenant. The skills point at a client configuration, a voice profile, a source list, and verification notes. Only the schemas are here. Every value is yours to write.
It does not adapt across regions. Spelling, region-native sourcing, and story substitution for a second market are a separate discipline and are not included.
It does not include the curator workflow: the process by which a source list, a ban list, or a standards file gets edited, reviewed, and versioned over time.
It does not draft. Every skill reviews. Writing the piece is upstream of this repository and stays there.
It does not know your clients. No real client name, data point, or identifying anecdote appears anywhere in the tree, including the worked example, which is fictional.
Each gap is deliberate. The system that composes the skills is not here either.
Read the playbook
The principles these skills enforce, in about five hundred words: The AI playbook