The skills library

Three selected skills for useful work: produce a Word file, trace an amendment history, or test a reusable workflow. Each entry explains what to supply, what it needs, and how to check the result.

Start with the task

A skill packages reusable instructions and may include scripts or reference files. These are annotated links to original packages, with Case & Code usage guidance.

Read Prompts, skills, workflows, and agents and Build a reusable task pack.

Create and check Word documents

Document production

When to use it: You need a reviewable Word file with formatting, comments, or tracked changes—not just draft text.

What the method changes: The source selects a document-editing method, uses supporting scripts, and requires rendering the result for inspection. The useful addition is an executable document workflow, not a more elaborate drafting instruction.

Original source

Read the original skill ↗
anthropics/skills · skills/docx/SKILL.md

Inputs: The original .docx file, the exact requested changes, an author name for tracked edits, and the required deliverables. Supply your own template if its formatting matters.

Tools and dependencies: A compatible agent with file and code execution access; the package and its helper scripts; document libraries and rendering tools such as LibreOffice and Poppler.

How to use it

Make the complete package available using your host’s skill-loading instructions. Supply the document and requested changes, then ask for a redline and clean copy.

Case & Code usage instruction

Using the available DOCX skill, edit [input.docx] only as follows: [changes]. Preserve all other wording and formatting. Attribute tracked edits to [author]. Return a redline and a clean copy. Validate the files and render every page for inspection. If a requested change cannot be tracked or the output cannot be rendered, stop and report the limitation before calling the task complete.

A first test

Use a fictional two-page agreement with a table, header, comment, and existing tracked deletion. Change one amount and one sentence. Check the redline in Word, confirm every other word is unchanged, and inspect both rendered pages. XML validation alone does not establish visual or substantive correctness.

Related lesson · Build a small test set

Back to the list

Trace a contract through its amendments

Document analysis

When to use it: A provision has changed across a base agreement and several amendments, and a two-document comparison no longer answers the question.

What the method changes: The original offers an amendment summary or a provision-by-provision trace, retaining document and section references. Our suggested adaptation separates execution dates from effective dates and keeps uncertainty about priority visible.

Original source

Read the original skill ↗
anthropics/claude-for-legal · commercial-legal/skills/amendment-history/SKILL.md

Inputs: The base agreement, every available amendment and schedule, stable document IDs, the provision or question, and the date as of which you need the analysis. Identify unsigned, disputed, or missing documents.

Tools and dependencies: A compatible skill-enabled agent able to read the supplied files. The original assumes practice instructions and may use a matter workspace. Direct file input is supported; its repository integrations are described as forthcoming.

How to use it

Inspect the source skill and its practice-file dependencies before loading it. Supply a bounded document set. Use the instruction below to narrow the task; it is Case & Code’s adaptation, not the upstream skill text.

Case & Code usage instruction

Trace [provision/topic] as of [date], using only [document IDs]. First inventory execution dates, effective dates, signatures/status, and missing referenced documents separately. Quote the original wording and each relevant change with document and section/page references. Check changes to definitions and cross-references. Distinguish an express textual replacement from your interpretation of its effect. If effectiveness, priority, or completeness is unresolved, retain the competing readings and identify what a lawyer must decide; do not declare a single controlling clause. Stop after the trace and unresolved-items list.

A first test

Use a fictional base agreement and three amendments: one signed later but effective earlier, one changing a defined term, and one referencing an absent schedule. Before running it, write the expected trace. Check all quotations, the indirect definition change, and whether the missing schedule remains unresolved.

Related lesson · Build a small test set

Back to the list

Build and evaluate a reusable skill

Workflow design

When to use it: You have a recurring task and want to know whether a reusable skill improves it over an ordinary instruction.

What the method changes: The source combines skill drafting with test cases, baseline comparisons, feedback, and activation tests. The transferable method is to assess both the work product and whether the skill is selected for appropriate requests.

Original source

Read the original skill ↗
anthropics/skills · skills/skill-creator/SKILL.md

Inputs: A narrow task, sample inputs, approved examples, explicit success criteria, prohibited errors, and requests where the skill should and should not activate. Reserve fresh cases for the final check.

Tools and dependencies: A compatible execution environment. Parts of the original use scripts, separate runs or subagents, and a Claude CLI-based activation-evaluation loop. Other hosts need an equivalent test harness; the file format alone does not provide one.

How to use it

Read the upstream workflow and dependencies. Start with a small skill. Hold inputs, model, tools, and source access constant when comparing it with a baseline. Separate gains from the instructions from gains due to extra tools, examples, or repeated calls.

Case & Code usage instruction

Use [available skill-creation workflow] to develop a skill for [task]. Inputs: [examples]. Success criteria: [criteria]. Prohibited errors: [errors]. Before running anything, propose a baseline, a version using the skill, and a version omitting [distinctive instruction]. Use the same inputs and available tools. Include [difficult case] and requests that should not trigger the skill. Record exact outputs, model/settings, calls, and who assessed them. If this environment cannot run separate comparisons, stop with the test plan and label it unexecuted. Do not invent scores or outputs.

A first test

Use a fictional date-extraction task with conditional and disputed dates. Set the answer key first. Compare missed dates, unsupported dates, preserved conditions, and source references. Test one near-miss request that should not activate the skill, then try a held-out document.

Related lesson · Build a small test set

Back to the list