Case study · AI systems
Claude Content Skills
An agency writes press releases and brand descriptions across fifteen-plus verticals. Each one carries rules that don't agree with each other — a newswire's editorial standard, State Bar advertising ethics, claim limits for regulated health categories, character ceilings that differ by platform. Holding all of that in one person's head, per client, every time, is the job these skills replaced.
Confidential mockup — all data fabricated
Six Claude skills, each a written instruction file plus reference sub-files — taxonomies, banned-word lists, category profiles. A skill selects itself from the shape of the request, applies the rule set that belongs to that vertical, runs the checks that used to be done by eye, and stops to ask when something doesn't verify.
The firm, attorney, case number, URLs and quoted copy in the walkthrough below are invented. One skill built for a single named client is shown under a generic name. The rules on display are the real ones.
01 — What the library holds
The legal and non-legal skills are deliberately separate files rather than one file with a flag. Legal content has to satisfy State Bar advertising ethics on top of the newswire's rules; non-legal content must not inherit them. Splitting them costs some duplicated formatting logic and removes the failure where one rule set leaks into the other.
02 — What's inside a skill
A skill is a written file, not a chat prompt typed fresh each time. Below is the one the walkthrough runs, trimmed to the four rules you're about to watch fire. The description block at the top is what makes the skill select itself.
---
name: press-release-creator-legal
description: Writes an EIN Presswire-compliant press release for a law
firm — general firm announcements and filed-case/case-result releases —
then saves it to Google Drive as a formatted Google Doc… Trigger even if
the user just pastes the intake block with no instructions, since that
terse format is exactly how these requests arrive.
---
## Step 3: Case-mode pre-draft gate
In case mode, verify all four of these before writing a single line. If any
fails, stop and flag it to the user — do not draft and then caveat.
1. The case is filed and verifiable. An anticipated, planned, or
threatened suit is a stop.
2. A court case/reference number is present, and it will appear in
the body.
3. Every factual claim traces to court documents.
4. Nothing implies a person committed a crime, lacks mental capacity,
or engaged in immoral behavior — unless citing a final court verdict.
This is the libel line, and it is easy to cross by summarizing
allegations as though they were established.
### Banned words — every one of these, in any form, anywhere
expert, expertise, specialist, specialize, best, indispensable, invaluable,
guarantee, ensure, ensuring, unique insight, navigate, navigating, serve as,
crucial, vital, instrumental, complexities, intricacies, meticulous,
landscape, paramount, firstly…
Substring matching catches most of these ("expert" catches "expertise",
"navigat" catches both forms). Run the check as a script, not by rereading.
### The boilerplate conflict — do not resolve this silently
The boilerplate is client-approved copy and gets inserted word-for-word.
Some approved boilerplates on this roster contain banned words. When the
scan hits inside the boilerplate: keep the boilerplate intact, and flag
the hit to the user. The user decides whether to edit approved client
copy. Never quietly rewrite a boilerplate, and never report a clean scan
when the boilerplate contains a hit.
Two of those rules exist because the obvious behaviour is wrong. A model that drafts first and caveats after has already produced something publishable. A model that tidies a banned word out of client-approved copy has edited a document the client signed off on.
03 — One slash command, start to finish
A real request arrives as a pasted intake block and nothing else. Here the news mentions a verdict, so the skill enters case mode — and the first thing it does is refuse to draft.
- 01 Invoke
- 02 Intake
- 03 Gate
- 04 Answer
- 05 Checks
- 06 Draft
- 07 Handover
The whole run, start to finish. A pasted intake block, a compliance gate that refuses to draft, and a checked release at the end.
Composer
Skills — 6 available
Pasted intake
press-release-creator-legal
Case mode. The news describes a verdict, so this is a filed-case release even though the intake didn't say so — the compliance obligations differ, and the pre-draft gate applies.
Gate 2 fails: no court case or reference number. It has to appear in the body, and it isn't in the intake. I'm not drafting around it and I'm not inventing one — a fabricated case number is a publishable-liability problem, not a formatting slip.
Also confirming before I write: is the judgment entered, and does every factual claim in the news block trace to a filing or the verdict form? Anything that only exists in the client's summary stays out.
No draft produced. Send the case number and I'll write it.
Reply
Checks — run as a script, not by rereading
- Body word count 446 — inside the 400–500 window
- Headline 69 chars, keywords inside the first 60
- Subtitle 115 chars, one sentence, present tense
- Dateline is City, State + em dash, carries no date
- Anchor hyperlinked on first instance only — 1 link in the body
- URL pasted byte-identical, UTM parameters unmodified
- Case number appears in the body; claims attributed to filings
- Banned word “expertise” found inside the client-approved boilerplate — kept word-for-word, flagged for your call, not rewritten
- Disclaimer required for TX contains “guarantee”, which is on the banned list — the mandated line overrides the style rule and stays verbatim
Output
Tarrant County Jury Awards $4.2 Million in Fort Worth Truck Collision
Ashgrove & Pell Injury Law represents the plaintiff in a commercial vehicle case tried in the 141st District Court.
Fort Worth, Texas — A Tarrant County jury has awarded $4.2 million to a Fort Worth motorist injured in a 2023 collision with a tractor-trailer on Interstate 35W, according to the judgment entered in the 141st District Court. Ashgrove & Pell Injury Law, a firm that represents Fort Worth truck accident lawyers clients across North Texas, tried the case over two weeks before the court entered judgment in July.
According to the complaint, the plaintiff was traveling southbound in the right lane of Interstate 35W when a tractor-trailer changed lanes and struck the driver's side of his vehicle. Court filings state that he was transported to a Fort Worth hospital on the day of the collision and later underwent two surgeries, one to his left shoulder and one to his lumbar spine. As alleged in court filings, the driver had exceeded federal hours-of-service limits before the collision, and the carrier's own electronic logging records were produced during discovery. A commercial vehicle safety analyst retained by the plaintiff reviewed those records and testified at trial, according to the court's docket entries. The defendants denied liability throughout the proceedings, according to filings in the case. The jury apportioned liability between the carrier and its driver and returned an award covering past and future medical care, lost earning capacity, and physical pain, according to the verdict form. The verdict form itemized the award by category rather than returning a single figure, and the court entered judgment on it the following week. Post-trial motions remain available to the parties under Texas procedure, and the judgment is subject to any appeal the defendants elect to file.
“The jury reviewed the records, heard the testimony, and reached its own conclusion,” said Renata Pell, Partner at Ashgrove & Pell Injury Law. “Our work was to present what the filings supported and let the evidence carry the case. A family managing a commercial vehicle collision is managing costs it did not choose, and that is the reason this firm handles these matters.”
The case was filed in the 141st District Court of Tarrant County, Texas, as Case No. 141-338291-25. Ashgrove & Pell Injury Law handles motor vehicle collision, commercial trucking, and premises liability matters for clients across North Texas, and represents plaintiffs in both state and federal court. The firm accepts personal injury matters on a contingency fee basis and offers a free initial consultation. The firm also handles wrongful death matters arising from commercial vehicle collisions, and its attorneys are licensed in Texas and have tried cases in courts throughout the region. Additional information about the firm's commercial trucking practice is available on its website, and the firm maintains offices in Fort Worth.
About — client-approved boilerplate, word-for-word
Founded in 2011, Ashgrove & Pell Injury Law represents injured Texans in personal injury and wrongful death matters. The firm's attorneys bring decades of combined trial expertise to catastrophic injury cases and have recovered compensation for clients across Tarrant, Dallas and Denton counties.
Disclaimer — required in TX
Attorney Advertising. Prior results do not guarantee a similar outcome.
Delivery message
Body 446 words. Saved to Drive, link above.
The quote is drafted, not real — the intake said SUGGEST ONE. It states a stance rather than a fact about the case, and needs attorney approval before publication.
Boilerplate carries “expertise”. Left exactly as approved. Your call whether to go back to the client or publish as-is — I'm not editing copy they signed off on.
Texas disclaimer added because the dateline state is TX. It contains “guarantee”; a mandated line overrides the banned-word list, so it's reproduced verbatim.
Transcript is a mockup. The rules, the gate and the checks are from the real skill.
The interesting output here is the third block, where nothing gets written. A skill that produces a good draft from a bad intake is the failure mode — the draft looks finished, and the missing case number surfaces at publication instead of at the keyboard.
04 — What the rules taught
Split the rule sets
Legal content answers to State Bar advertising ethics as well as the newswire; non-legal content must not inherit those restrictions. One file with a flag would eventually apply the wrong set, so legal and non-legal are separate skills — duplicated formatting logic accepted as the cost.
Flag approved copy, don't fix it
Client-approved boilerplate sometimes contains a banned word. Silently correcting it means editing a document the client signed off on. The scan keeps the boilerplate intact, names the word and the sentence, and hands the decision back — and never reports a clean scan when there was a hit.
Stop, don't draft and caveat
Every verification gate is checked before the first line is written, not after. High-risk claims that can't be confirmed are left out rather than softened with a hedge, because a hedged claim in a finished draft still reads as publishable.