Let me look at a couple of examples for house style.Now I understand the format. Let me check the environment's Node version to decide how the example's TypeScript runs offline.The setup is clear now. Node 22.23 has type-stripping enabled by default, so a zero-dependency TypeScript example can run and be tested with no installs — ideal for an offline verifier.

Decision up front: this is a runnable skill (role: implement), not a reference. The task only makes sense as "do this, and here's how we know you did" — it demands a delivered bot, acceptance criteria, a write budget, and a judge. It'll be checks: none + verifier-paths since no verifier suite installs npm packages and runs a TypeScript harness.

Let me plan and write the files.Now the worked example — a zero-dependency TypeScript Telegram bot that runs under Node's built-in type stripping, so the harness needs no token, no network, and no installs. Careful: type stripping requires erasable-only syntax (no constructor parameter properties, import type for interfaces).Now the source files.Now the README section listing it, plus a row in the "What is here" table.Checker passes and all 8 harness tests pass. Let me verify --help shows the notice and do a final self-review of the SKILL.md for subtle checker rules.One nit: the criterion requires the notice "verbatim at the top of the README" — in the example README I wrapped it across two lines, which a literal-minded check would miss. Make it one line.Done. node check-skill.mjs build-chat-bot reports ok and the example's own harness passes 8/8 with no token, no network, no installs.

Decision

Runnable implement, not a reference. The task only makes sense as "do this, and here is how we will know you did" — it demands a delivered bot, criteria, a budget and a judge. Nothing in it would still be true with no job in front of it, which is the guide's test. It's checks: none + judge: verifier-paths (class 2): no suite in the verifier's closed set installs npm packages and runs a TypeScript harness, so claiming verifier-rerun would pretend a suite exists that doesn't. writes: any, because the work is a new project whose layout the request decides — no fixed path list can name it in advance. Each frontmatter field carries a # comment justifying it.

What I wrote

• build-chat-bot/SKILL.md — objective passes {{objective}} through; five acceptance criteria a person can check from the delivered tree (exact command set, documented harness passing token-free, env-only secrets with env.example, the experimental notice verbatim on README top/--help/every served page, followable deploy notes). Body is one screen: first moves, bold-led rules, what failure looks like.
• build-chat-bot/REFERENCE.md — platform transport notes (Telegram long-poll vs webhook, Discord Gateway vs Interactions endpoint), secrets hygiene, the fake-transport harness pattern, rate-limiting defaults, deploy-note content. Reached via reads: [skill:build-chat-bot].
• build-chat-bot/example/ — a zero-dependency Telegram bot that is the smallest output satisfying the criteria: src/ (env-validating entrypoint, command registry, injected-clock rate limiter, Transport interface with live TelegramTransport and ScriptedTransport, status server carrying the notice), test/bot.test.ts (8 node --test cases covering every command, the limiter including window reset, the allowlist, and env loading), env.example, Dockerfile, and a README ending in systemd + container deploy notes. It runs on Node ≥22.18's built-in type stripping, so nothing needs installing — I avoided non-erasable TS (e.g. constructor parameter properties) deliberately.
• README.md — a row in "What is here" and a short section describing the skill and its example, quoting the experimental label.

Notes

• The label appears verbatim at example/README.md top, in --help output, in package.json description, and on the status page. I did not

Recorded 7 findings: 6 medium, 1 low in .imd-findings.json.

Issues include missing verifier-enforced deliverables, unverifiable command completeness, an allowlist that fails open, command lookup errors, duplicate replies after retries, missing outbound pacing, and README labeling.

Every finding includes concrete reproduction details and verified source locations.