Contents

Cheatsheet · Claude Code

Everything this course asks you to write in Claude Code, collected on one page. This is a lookup sheet for people who have already worked through the steps — it repeats shapes, not explanations. For why any of it exists, go back to the step or the kit it came from.

Where files go

Path	What lives there
`.claude/skills/<skill-name>/SKILL.md`	One skill. The directory name is the skill name.
`.claude/agents/<agent-name>.md`	One subagent, as a single file.
`schema.example.json`	The kit's fixed type list, at the kit root.
`sample-input.md`	The document or record a kit's skill reads by default.

Claude Code picks both up from the path alone. There is no separate registration file, and nothing has to be listed anywhere else before it can be invoked.

Skill front matter

yaml

---
name: extract-facts
description: Extracts Service/Incident/Cause facts from an incident-report document against a fixed schema, rejecting anything the schema doesn't define
model: claude-opus-4-1-20250805
temperature: 0
tools: [Read, Write]
---

Field	Notes
`name`	Matches the containing directory.
`description`	One sentence, written so a reader can tell what gets rejected, not just what gets produced.
`model`	Pin it when a kit's behavior depends on the model; omit to inherit.
`temperature`	Every kit in this course sets `0`. Extraction and checking are not places you want sampling variety.
`tools`	Least privilege. An extraction skill gets `[Read, Write]`; a verifier gets `[Read]` only, so it cannot repair what it is supposed to be judging.

Skill body

The kits use the same five headings in the same order:

Heading	Holds
`## Instructions`	Numbered steps, in execution order. Step 1 is nearly always "read the schema before you read anything else."
`## Input`	Each file the skill opens, and what it is for.
`## Output`	Each artifact written, plus anything printed for a human.
`## Example Usage`	A concrete run with its expected result — including the expected drops.
`## Validation`	Which agent or mode re-checks this skill's output.

Two habits carry most of the weight across every kit:

Freeze the schema at step 1. Say so in the instructions. A skill told to read the type list first, and to treat it as fixed for the rest of the run, is much harder to talk into inventing a type mid-extraction.
Make the drop list an output, not a side effect. A run that reports only what it kept cannot be reviewed. Every extraction kit here prints what it rejected and which check each item failed.

Subagent front matter

yaml

---
name: graph-verifier
context: verify-graph-output
tools: [Read]
---

Body headings used by the kits: ## Purpose, ## Inputs, ## Validation Steps, ## Output, ## Example.

A verifier in this course is written to do an independent pass, not to read the producing skill's self-report and agree with it. Three rules make that real:

It re-derives the allowed set from the schema itself.
It reads the output file directly, rather than being handed a summary.
It cross-checks the producing skill's claimed drop list against what it actually found, and flags any disagreement as its own finding — a mismatch is evidence the run's self-report cannot be trusted.

Give it tools: [Read]. A verifier that can write can quietly fix what it should be reporting.

Verdict shape

Checkers in this course report a decision plus the single piece of evidence it rested on:

text

PASS/FAIL: FAIL

- affected: INC-4482 -> "Priya Raman" — FAIL: endpoint type is Person,
  not Service. "affected" must run Incident -> Service.
  Fix: drop this relationship; Person is not a schema entity type.

Name the rule broken and the fix. A bare FAIL sends the reader back to re-run the check by hand to find out what happened.

Invoking

Claude Code invocation is conversational — ask for the skill by name and name the files it should work on:

text

Run the extract-facts skill on sample-input.md using schema.example.json.

Then run the verifier as its own turn. Keeping the two apart is the point: a producer asked to check itself in the same breath tends to grade its own work generously.

Kits with a Claude Code implementation

All 23. The seven core kits ship a Claude Code implementation and an OpenCode one; the other sixteen ship Claude Code plus a PORTING.md. See starters/README.md for the split, and opencode.md for the other half of the pair.