feat: add skill-creator casedev skill

skills/casedev/skill-creator/SKILL.md

@@ -0,0 +1,97 @@
+---
+name: skill-creator
+language: en
+description: Create a new private skill in the case.dev skills registry for the user's organization. Use when the user wants to create, save, build, or register a new skill or reusable workflow — e.g. "create a skill", "save this as a skill", "turn this process into a workflow", "make this reusable". Covers interviewing the user, drafting the skill, and creating it with the casedev CLI.
+---
+
+# Skill Creator
+
+Create a new **private (org-scoped) skill** in the case.dev skills registry. A
+skill is a reusable instruction set a legal AI agent loads with `skill_read`
+and follows exactly. Once created, the skill appears in the organization's
+Skills page and can be attached to any chat with `$`.
+
+## Process
+
+Work through these four steps in order. Never create the skill before the
+user approves the draft.
+
+### 1. Interview
+
+Ask only what you need to draft well — usually one short round of questions:
+
+- What should the skill do? (the task, end to end)
+- What does it work from? (matter documents, user-provided facts, both)
+- What output is expected? (document type, structure, length, tone)
+- Any firm-specific conventions? (templates, phrasing, review steps,
+  jurisdiction defaults)
+
+If the user already described the task in detail, skip questions you can
+answer from their message.
+
+### 2. Draft
+
+Draft all three fields and show them to the user for approval:
+
+- **name** — short, imperative, title case. `Demand Letter Draft`, not
+  `A skill that helps with demand letters`.
+- **summary** — one sentence saying what the skill does **and when to use
+  it**. Concrete trigger phrases beat vague descriptions: agents decide
+  whether to load a skill from this line alone, and they under-trigger on
+  vague summaries. Good: "Drafts a demand letter from the matter record;
+  use when the user asks for a demand letter, settlement demand, or
+  pre-suit letter." Bad: "Helps with letters."
+- **content** — markdown instructions written **for the agent, not the
+  user**: step-by-step procedure, what inputs to gather, exact output
+  format, and quality checks before delivering. State the steps as
+  commands ("Extract every treatment date", not "Treatment dates could be
+  extracted").
+
+### 3. Create
+
+After explicit approval, create the skill with the pre-authenticated casedev
+CLI. Write the JSON to a file first to avoid shell-quoting issues:
+
+```bash
+cat > /tmp/skill.json <<'EOF'
+{
+  "name": "Demand Letter Draft",
+  "summary": "Drafts a demand letter from the matter record; use when the user asks for a demand letter, settlement demand, or pre-suit letter.",
+  "content": "# Demand Letter Draft\n\n..."
+}
+EOF
+casedev api POST /skills --body "$(cat /tmp/skill.json)"
+```
+
+The response includes the created skill's `slug`. Skills created this way are
+**private to the organization** — they are org custom skills, not published
+anywhere public.
+
+To update an existing skill, use `casedev api PUT /skills/<slug>` with the
+same body shape. To verify a skill exists, use the native `skill_read <slug>`
+tool.
+
+### 4. Confirm
+
+Report the result in plain language: the skill's name and slug, that it now
+appears under **Skills → Your firm**, and that it can be attached in any chat
+by typing `$`. Suggest a quick test run.
+
+## Rules
+
+- Never create a skill before the user approves the full draft.
+- Never paste raw API responses at the user — report results in plain
+  language.
+- Never print, echo, or reveal `CASEDEV_API_KEY` or any credential.
+- The CLI is pre-authenticated: never run `casedev login` and never pass
+  `--api-key`.
+
+## Troubleshooting
+
+- **409 / slug conflict** — a skill with that name already exists. Ask the
+  user whether to update the existing skill (`PUT /skills/<slug>`) or pick a
+  different name.
+- **401/403** — the runtime key cannot write skills; tell the user the
+  creation failed and to contact support. Do not retry in a loop.
+- **Content too long** — skill content is capped (64KB). Trim reference
+  material; link to matter documents instead of inlining them.