Teach this: the Writing Partner

Sign in through Stanford, not with a personal account. The university versions are the ones covered by Stanford’s agreements and its guidance on what may be put into them; a personal free account is not, and what you paste into one is not governed by the same terms.

The point of installing in both tools and as an agent is not thoroughness for its own sake. It is that the three set-ups behave differently, and the differences are most of what there is to discuss. Ask participants to finish all three before the session; budget twenty minutes at the start if you would rather do it together.

Everything below comes from one download: the SLS Writing Partner Set. Unzip that set once. Inside are ten skill files, each its own ZIP, and each of those stays zipped at every step that follows.

Five fictional student drafts sit on the skills page under Practice drafts: a case brief, three memos of increasing length and citation density, and a timed exam answer. They were written as teaching samples, and each one has real weaknesses built into it — that is what makes them useful. Nobody has to volunteer their own writing to have something worth reviewing in front of the room.

Which draft to hand out depends on what you want the room to find:

The question: what is a skill, and why does that matter to a lawyer?

You need: the case study, Reverse-engineering Anthropic’s AI governance legal skills. Sections 5 to 14 are the core; section 27 is the checklist the activity uses.

1. Read sections 5–14 in advance, or read section 6 and 7 aloud together (five minutes).
2. In pairs, open any one of the ten writing skills you installed and answer four of the twelve forensic questions from section 27: what triggers it, what single job it owns, what it is not authorised to do, and who takes over next.
3. Each pair reports the boundary they found — the thing their skill refuses to do.

Watch for: the moment someone realises the file is just written instructions, not a separate intelligence. That realisation is what makes the rest of the session possible.

Debrief: If a skill is a written procedure, whose procedure should it be — Anthropic’s, a vendor’s, the library’s, or your professor’s? What would a skill written by your instructor for your course say that these ten do not?

The question: how much of what comes back is the skill, and how much is the model running it?

You need: one practice draft, the same skill installed in ChatGPT and in Claude, and an identical opening message in both. The Morgan memo with Claims and Source Traceability is the sharpest pairing; Bluebook Audit is a close second.

1. Send the same file and the same request to both tools. Do not help either one.
2. Put the two outputs side by side and mark, for each: what both found, what only one found, and anything either asserted that the other contradicted.
3. Then check three of the findings against the draft itself. Which tool was right?

Watch for: confident disagreement. Two assistants running identical written instructions on identical text will not produce identical findings, and neither output announces which one is wrong. Watch also for the tempting conclusion — that the longer review is the better one.

Debrief: If two runs of the same skill disagree, what does that tell you about a single run you never compared? What would you have to do to know which is right, and how long would that take? Is the answer “use both every time” realistic?

The question: what does a full review actually give you, and what does accepting it uncritically cost?

You need: the ChatGPT agent from set-up C, and the Morgan or Lee memo.

1. Hand the agent the draft and ask for a full review, including the Word review copy.
2. Read the Human Review Required section it produces before reading anything else. Does it name what you would have named?
3. Take three findings and rule on each: accept, decline, or revise differently — and write one sentence of reasoning for each ruling. Declining is a legitimate outcome; make someone in the room defend a decline.
4. Take one flagged authority and try to verify it properly. Notice how long it takes.

Watch for: the pull towards acceptance. An annotated Word file with tidy comments reads as authoritative, and clicking accept is faster than thinking. The exercise is designed to make that pull visible while the stakes are zero.

Debrief: Whose writing is it after you accept twenty suggestions you did not evaluate? Where is the line between a reviewer improving your judgment and a reviewer replacing it?

The question: where is the boundary, and how firm is it?

Every one of these skills refuses the same thing: it will not draft, rewrite, or hand back replacement prose. This activity tests that claim rather than trusting it.

1. Give a draft to any focused check and try, in three or four turns, to get it to write the fix for you. Ask for “an example of how that sentence could read.” Ask it to “show, not tell.” Say you are out of time.
2. Record what it refused, what it conceded, and the exact wording that came closest to working.
3. Compare across the room. The same pressure will not produce the same result twice.

Watch for: the grey zone. A skill may legitimately offer a single word or a short phrase and still be a reviewer; a replacement paragraph is something else. Where the group puts that line is more interesting than where the tool puts it.

Debrief: If you can talk a tool past its own boundary in four turns, what is the boundary actually for — and what is left holding the line? (Answer: the Honor Code, the course rule, and the person typing.) Would you disclose the turn that worked?

The question: can you encode your own standard well enough that a machine could follow it?

You need: the case study’s teaching template (there is a copy button on it), and a real standard from a course, journal, or clinic — a rubric, a style rule, a citation convention, a professor’s standing complaint.

1. Copy the template and fill in the trigger and the description first: when should this run, and how would the assistant know?
2. Write the workflow as numbered steps, then the output contract, then the boundary — what it must not do.
3. Trade drafts with another group and run the twelve-question checklist against theirs.

Watch for: how quickly “write well” turns out to be unwritable, and how much of legal writing instruction is tacit. Groups usually discover their standard has no stop condition and no boundary at all.

Debrief: What part of your standard could not be written down? Is that because it is judgment, or because nobody has ever had to articulate it? Which answer should worry a teacher more?

• If a reviewer tells you a paragraph is unclear and you rewrite it, whose paragraph is it?
• What if the reviewer names the exact problem and you make the obvious fix?
• Is a human writing tutor doing something different in kind, or only in degree?
• At what point would you disclose the help, and to whom?
• Does it change if the reviewer is a classmate, a TA, a paid editor, or a skill?
• Would your answer change if the work were a journal note rather than a memo?

Listen for: the slide from “it only diagnosed” to “so the revision is entirely mine.” Both halves can be true and the conclusion can still be too comfortable. Push on the case where the diagnosis was the insight.

• What is the difference between a source that exists and a source that supports the point?
• How would you know if a quotation had been subtly altered?
• When does a public-web check stop being enough?
• What does a citator tell you that no assistant can?
• How long did verifying one authority take in activity C?
• What happens to that discipline at 2am the night before?

Listen for: people treating “the assistant found the case” as the end of the process. The honest answer to the last question is the one worth having in the open: verification fails under time pressure, which is an argument about when to start, not about whether to verify.

• Does installing a skill make the assistant smarter, or only more specific?
• Why ten small skills instead of one large one?
• What could a skill never fix about the underlying model?
• If a firm encodes its house method as skills, what has it actually built?
• Who should be allowed to write the skills a law school distributes?
• What would you want to know before installing one from a stranger?

Listen for: the assumption that a skill is a guarantee. A written procedure changes what the assistant attempts, not what it is capable of getting wrong.

• Where did the two assistants differ in activity B, and did it matter?
• What does a legal platform give you that a skill cannot?
• What does an open skill give you that a platform cannot?
• Which of these will you have access to after graduation?
• Whose data governance applies to what you paste in?
• What should a firm ask before buying either?

Listen for: winner-takes-all framing. The case study’s stack diagram is the corrective: the interesting question is which layer you are choosing at, not which company wins.

• Which assignments should get no AI review at all, even where permitted?
• What are you practising when you find your own weak paragraph?
• What does a first-year lose by outsourcing the diagnosis?
• What would you never put into any of these tools, and why?
• How would you word an AI-use disclosure you would be comfortable signing?
• What would make you stop mid-review and close the window?

Listen for: the difference between prohibited and unwise. Both matter, and only one of them is written down. The PAUSE Rule is the site’s workflow for the second.

SLS AI Use Gate

Decides whether AI review of this draft is authorised at all, under the course rule or the SLS default, and flags confidentiality and professional-responsibility problems before anything reads the paper. Say out loud: this is the only skill in the set whose answer can be “stop.” A group that skips it has already made a decision without noticing.

SLS Writing Review

The umbrella workflow: the gate, then argument, organisation, clarity, audience, counterarguments, sources, genre, and Bluebook, ending in an annotated Word copy. Say out loud: it hands work to the nine others, so a set missing three of them quietly reviews less than it claims.

SLS Argument and Structure

Tests the thesis, rule synthesis, evidence chain, and conclusion for gaps, then asks the revision question rather than answering it. Say out loud: the questions it asks are the ones a supervising attorney asks. Reading them is training even when the diagnosis is wrong.

SLS Flow and Organization

Sequencing, headings, transitions, repetition, and the point buried where no reader will reach it. Say out loud: the easiest skill to agree with and the easiest to over-apply — a perfectly signposted argument can still be wrong.

SLS Clarity and Precision

Ambiguity, jargon, dense sentences, inconsistent defined terms, flagged in place with revision strategies rather than replacement sentences. Say out loud: this is where the reviewer/ghostwriter line gets tested hardest, because a “clearer version” is one request away. Activity D lives here.

SLS Audience and Reception

How a professor, judge, supervising lawyer, client, editor, or employer is likely to receive the draft: tone, credibility, framing, unstated assumptions. Say out loud: it is guessing at a reader it has never met. Its value is in making you specify the reader, which most drafts never do.

SLS Counterargument Stress Test

The strongest objections, missing adverse authority, doctrinal soft spots — issue-spotting and questions, never a rebuttal. Say out loud: the most useful of the ten for a strong writer, and the one whose output you are least able to check without doing the reading yourself.

SLS Claims and Source Traceability

Traces each claim to its source and separates “this source exists” from “this source supports the point,” ending in a claim–source matrix. Say out loud: the distinction it draws is the one that ends careers when it collapses. Pair it with the Morgan memo and make someone open an authority.

SLS Bluebook Audit

Bluebook 22nd-edition issue-spotting: short forms, pincites, signals, order, typography. It names the likely defect and the rule; it does not write the citation. Say out loud: it marks HUMAN VERIFICATION REQUIRED where the rule is uncertain or the source was never opened — and a confident citation correction from an assistant that never opened the source is exactly the failure mode to fear.

SLS Genre Fit

Whether the draft behaves like the form it claims to be — memo, brief, seminar paper, client letter — on purpose, stance, sections, and conventions. Say out loud: the best opener for a 1L room, because genre confusion is the most common and least-discussed failure in first-year writing.

Reverse-engineering Anthropic’s AI governance legal skills takes apart a real, public, Apache-2.0 project — Anthropic’s Claude for Legal — and asks what a skill is made of. It is the reading that turns “I installed a thing” into “I know what I installed.”

There is also a 21-minute audio explainer covering the same ground. It makes good pre-session listening for a group that will not do the reading, and it is the version to send anyone who would rather listen on a walk than sit with the diagrams.

Sections 1–4 · skill or platform

Why an open legal skill does not put Harvey or Legora out of business, and what the layers of the legal AI stack actually are. Use it for: discussion set four.

Sections 5–14 · anatomy

What a SKILL.md file contains, what YAML frontmatter is, and why the description is what routes work to the right skill. Use it for: activity A, and for anyone who thinks a skill is code.

Sections 15–23 · design lessons

Small skills over one large one; the file as a control panel rather than a warehouse; what belongs in references, scripts, and assets instead. Use it for: activity E.

Sections 24–28 · the checklist

Twelve questions to ask of any skill, ending with the one that matters most: what does this depend on that the file itself does not contain? Use it for: evaluating anything anyone hands you.

Nothing here is fixed. Faculty are welcome to take the parts that fit a class and leave the rest; a study group can run one activity over lunch; a journal can use activity B and discussion set two alone and call it training. Three ways it bends:

• Swap the material. Any practice draft works with any skill; the table above is a suggestion, not a pairing rule.
• Swap the tool. Where a group has one platform, run everything there and demonstrate the other.
• Add your standard. Attach your own rubric or style guide to the agent and the review changes with it.
• Narrow it. A single skill, a single draft, and one discussion set is a legitimate twenty-minute session.
• Print it. This page is written to print cleanly as a handout.
• Ask us. The library will run it with your class if you would rather not.