Nagi 凪

Start messy. Nagi's compiler is designed to extract signal from noise. Don't overthink your input — paste the raw thought, error log, Slack thread, or half-formed idea. The compiler will structure it.

Good: "Login form submits but shows blank page instead of /dashboard"

Weak: "Fix the login"

Name the behavior you see, the behavior you expect, and where it happens. Nagi fills in the engineering structure.

When you mention files (src/components/Auth.tsx, styles/nav.css), Nagi detects them and routes them into the right context section. This dramatically improves the output.

Compound requests ("fix the login AND redesign the nav AND add tests") produce diluted output. Split them into focused prompts. Each transform should have one clear objective.

When the prompt strength bar shows below 60%, Nagi suggests a template. Templates aren't bureaucracy — they're the fields that make agents perform. Fill the starred fields at minimum.

Good: "Don't change the auth flow, just fix the redirect"

Constraints prevent agents from over-engineering. Tell them what NOT to do and they'll stay focused.

After compiling, use the refine chips (Deeper, Shorter, Code only, Ship-ready) to adjust the output. You can also add context and recompile without starting over.

Auto — Nagi detects the right mode from your input. Works 90% of the time.

Claude Code — Full directive with Explore/Plan/Implement/Verify structure and guardrails.

Refine — Clean up and restructure an existing prompt.

Compress — Strip noise from verbose context while keeping signal.

Audit — Find quality issues in a prompt and get a rewrite.

Personal Ops — Life admin: triage, scheduling, drafting messages.

Shorten — Maximum brevity while preserving intent.