Get started with K13

1. Drop in the skill

The fastest way to adopt K13 is a skill file. Download it, put it in one folder, and the agent reads it. No install, no keys, nothing to run.

mkdir -p .agents/skills/k13
curl -fsSL https://randomknights.xyz/source/K13-SKILL.md \
  -o .agents/skills/k13/SKILL.md

The layout is the whole configuration:

your-project/
  .agents/
    skills/
      k13/
        SKILL.md      <- the file you just downloaded

That folder is the Agent Skills format, so the same file works in the Codex CLI, in Claude Code and in the ChatGPT desktop app. It is self-contained: the whole standard is inline, so it needs no network and no second file once it is on disk.

The published skill is byte-locked like the standard itself. Its sha256 is in /source/source-lock.json, so you can pin the exact copy you adopted.

2. Or take the steps into your own process

If your pipeline is not agent-shaped, the standard is still the deliverable. Run these thirteen steps where the work happens, in whatever form your process already uses: a checklist, a template, a prompt.

1  Verify state. Real sources only. No guessing, no plausible defaults.
2  Capture the cost before drafting anything else.
3  State the outcome first, then the method. What happened, not effort.
4  Flag every roadblock. Blocked, skipped or partial, named plainly.
5  Translate the cost into meaning, not raw figures.
6  Name the owner action. Exact click, exact command.
7  Pull the next-command template. Model and effort included.
8  Check the humour. Sparse, honest, off entirely on safety, money or loss.
9  Draft the juice box pass. K-5. No jargon at all.
10 Draft the soda pop pass. 6-8. One new term at a time.
11 Draft the energy drink pass. 9-12. Real terms, defined once.
12 Draft the black coffee pass. College. Full precision.
13 Run the gate. No em dashes. Every claim checked. All four passes agree.

3. Pick a default level

Energy drink, grades 9 to 12, is the level a reader gets when no level is chosen. Set the level in the request itself so one honest answer is written at that level. Do not generate at one level and rewrite into another: that costs twice and lets the meaning drift from the original.

The demo shows one fact written at every level, so you can see what changes and what must not.

4. Keep one command template

Keep exactly one canonical, version-controlled command template and build every dispatch from it. Where that file lives is your business. The shape is the part the standard fixes:

task:
repo:
model:
effort:
scope:
constraints:
verify steps:
stop condition:

5. Gate what a machine can actually check

Five rules in K13 are mechanically checkable, and a validator can honestly enforce every one of them:

• No em dashes and no en dashes, in the response, code comments, docs, commit messages and public copy. One is enough to fail.
• ASCII in committed text.
• Level names spelled plainly, no numeric swaps.
• Every figure identical across levels. If two levels disagree on a number, the answer is wrong.
• A cost line present, or an explicit statement that usage could not be reported.

The dash rule is the cheapest to wire in, and it belongs in continuous integration:

# The dash rule is the one step a machine can fully check for you.
# Fails the build if a tracked text file carries an em dash or an en dash.
git ls-files -z | xargs -0 grep -nP '\x{2014}|\x{2013}' && exit 1 || exit 0

Everything else in K13 needs judgement, and no validator can honestly gate it. Whether a fifth grader could follow the juice box pass, whether the humour earns its place, whether a claim was actually checked: those are read by a person or they are not read at all. A conformance badge that implied otherwise would be measuring the five rules above and calling it the standard.

Cite what you built against

The standard is version 1.1.0, CC BY 4.0. Its exact bytes are at /source/K13.md, sha256 f4a5e8dae25a. Cite the version together with that hash, so a reader can tell which text you meant.