a knowledge pack for coding agents · from the s&box field guide
Give your agent the gotchas
before it hits them.
The Agent Kit is the s&box Field Guide, repackaged for Claude Code, Cursor, and any agentic coding tool: dense, per-lane checklists of every verified trap and fix, plus the full articles behind them, that your agent loads before it writes engine code.
880+
verified gotchas, split by lane
10
task lanes, set up → ship
✓ verified
every fix stamped with engine build
0
invented fixes: unverified claims get flagged
01
Pull it into your project
terminal
# clone the kit next to (or inside) your project
$ git clone https://github.com/S-box-Field-Guide/agent-kit
# keep it fresh: the corpus is curated upstream
$ cd agent-kit && git pull
Then point your agent at it
Reference the kit from your
CLAUDE.md / AGENTS.md, or install SKILL.md as a Claude Code skill. The agent takes it from there.Works with
Claude CodeCursorany agent that reads Markdown
02
What's in the kit
_core.md
Always-load core rules
The handful of cross-cutting rules that cause most failures: units are inches,
dotnet build doesn't enforce the whitelist, BuildHash gates every re-render. Tiny. The agent loads this first, every time.<lane>.md
Per-lane checklists
Ten scan-first packs, one per task lane. The agent loads only the one it needs (
getting-art-in, writing-gameplay, building-ui…) for a dense list of traps and their fixes. Full Symptom / Cause / Fix / Why articles link out from coverage.md.SKILL.md + index.md
The routing map
Tells the agent which lane pack to load for the task at hand. "Importing a Blender mesh? Load
getting-art-in.md." Plus the rules: never invent a fix, flag anything unverified. coverage.md tracks what's covered vs. the backlog.03
How your agent uses it
step 1
Load core, then the lane
Before touching engine code, the agent loads
_core.md, then the one lane pack for the task, and scans for matching bullets: units are inches, Blender +X faces world −Y, BuildHash is the only UI re-render trigger…step 2
Open the matching article
When a bullet matches, it reads the full Symptom / Cause / Fix / Why, and applies the documented fix instead of guessing at engine behavior.
step 3
Never invent, always flag
Anything the kit doesn't cover gets marked
(needs verification) rather than fabricated. Corrections go back upstream as GitHub issues.