Nimi Coding
Nimi Coding is a vendor-neutral AI-native methodology product for governing high-risk AI-assisted software work. It ships as a standalone npm package (@nimiplatform/nimi-coding), bootstraps a project-local .nimi/** truth surface into any repository, and turns "AI plausibly finished this" into "the four closure dimensions are evidenced."
Nimi Coding is one of the products inside the Nimi platform — the AI development methodology that ships with everything else. It can also be adopted on its own: the package is host-agnostic and works in any repository, regardless of whether the rest of the Nimi platform is in use.
Nimi Coding and the rest of the platform stress-test each other. Nimi Coding is what makes a system as ambitious as Nimi buildable by a small team using AI; the platform's actual scale is what makes Nimi Coding's claims falsifiable in practice.
Why This Section Exists
Most AI products solve "AI in the editor." Nimi Coding solves "how does anyone trust the work AI did?" The answer is not better prompts and not better tests. It is methodology — explicit machinery for declaring closure conditions before work begins, and verifying them as evidence after work ends.
If you have ever watched an AI-assisted change look complete to every available signal — type checker green, tests green, code review approved — and turn out to be wrong about authority, scope, or product meaning, this section is for you.
Start Here If You Are New
The first successful Nimi Coding path is intentionally small:
- Install the package in an existing repository. See Installation.
- Bootstrap
.nimi/. Usenimicoding start, then check the result withnimicoding doctor --json. - Reconstruct project authority into
.nimi/spec/**, recording source basis and unresolved gaps instead of inventing clean rules. - Create a topic for the first high-risk or authority-bearing change.
- Split the topic into waves so each wave has one owner domain and one closure goal.
- Freeze a packet before work starts: allowed reads, allowed writes, acceptance invariants, negative tests, stop lines, and reopen conditions.
- Run or hand off the work through an admitted AI host, then record typed evidence.
- Close the wave only across all four dimensions: authority, semantic, consumer, and drift resistance.
That path is the product in miniature: AI work becomes durable, bounded, auditable project state instead of a chat transcript that looked convincing at the time.
What's In This Section
The Paradigm
- The Paradigm — what's actually new about AI-coding governance and why this is a paradigm rather than a checklist.
- Four Closures — authority, semantic, consumer, and drift-resistance closure as a thinking framework.
- False Closure Typology — the named failure shapes the methodology catches.
- Forbidden Shortcuts — the catalog of refused anti-patterns.
Roles And Convergence
- Role Separation — manager, worker, auditor.
- Authority Convergence — why an independent audit has to come before implementation when spec changes.
Lifecycle
- Topic Lifecycle — proposal, ongoing, pending, closed; wave fine-grained states; true close.
- Whitepaper — the conceptual case argument for treating AI-assisted implementation as authority-bearing work.
- Topic Workflow — the operational topic / wave / packet / preflight / audit / closeout flow.
- Walkthrough — a synthetic example end-to-end.
The Package
- The Package — what
@nimiplatform/nimi-codingships, what it does not ship. - Host-Agnostic Boundary — why switching AI hosts does not change the methodology.
- Skills — the four declared skills (
spec_reconstruction,doc_spec_audit,audit_sweep,high_risk_execution). - CLI Surface — concept-level overview of the command surface.
- Installation — current installation posture.
Comparison And Adoption
- Comparison — vs vanilla AI coding, code review, DevOps governance, DDD, agile.
- Adoption Path — who would adopt this and why.
Practical Sub-Trees
- Tutorials — learning-oriented step-by-step lessons, including the full path from install to
.nimi/spec/**, topic execution, sweep audit, sweep design, and long-running host work. - How-to — problem-shaped recipes.
- Reference — schema-level dictionary.
Appendix
- oh-my-codex Adapter — admitted external host adapter overlay.