Skills

The Nimi Coding package admits four declared skills. Each skill is a typed surface that an admitted external AI host can fulfill. Skills are the formal contract through which the host performs Nimi-Coding-governed work.

The Four Skills

Skill	Required	Purpose
spec_reconstruction	yes	Reconstruct project canonical tree into .nimi/spec/**
doc_spec_audit	yes	Compare human-authored docs against .nimi/spec/**
audit_sweep	no	Perform full-coverage audit sweep and emit a frozen findings ledger
high_risk_execution	no	Packetized execution after project truth matures

spec_reconstruction (required)

The first skill any new project runs. Converts whatever a project currently has — mixed code, scattered docs, ADRs, README files — into a canonical authority tree under .nimi/spec/**.

Input	Output
Mixed inputs (code, docs, structure, human notes)	Canonical tree under .nimi/spec/** plus .nimi/spec/_meta/spec-generation-audit.yaml

Property	Value
Trigger	bootstrap_only
Output rule	"every generated canonical file requires explicit source basis and unresolved gap tracking"
Hard constraint	"no human-friendly parallel truth"

Reconstruction does not invent. Every generated rule has either explicit source basis or explicit unresolved-gap tracking.

doc_spec_audit (required)

After reconstruction, this skill checks human-authored docs against the canonical spec tree. Drift detection.

Input	Output
Human docs + canonical spec tree	Drift findings ledger

A doc that drifts from spec is a finding. A doc that re-states spec is OK. A doc that contradicts spec is a critical finding.

audit_sweep (optional)

Full-coverage audit sweep over the project. Emits a frozen findings ledger.

Input	Output
Project corpus	Findings ledger (frozen)

The "frozen" property is what makes the ledger usable as evidence. A sweep result is recorded; it does not get edited later.

high_risk_execution (optional)

Packetized execution after project truth matures. This is the skill the methodology was actually designed for: high-risk work that needs the four-closure framework.

Input	Output
Frozen execution packet	Worker output + evidence

A high_risk_execution skill run consumes a frozen packet and produces output that the closeout step can verify.

Property	Value
Trigger	After .nimi/spec/** is canonical-tree-ready
Owner	Manager (admits the run)
Auditor	Independent (per role separation)

How Skills Are Dispatched

The package's nimicoding handoff command emits an authoritative machine-readable handoff payload.

Field	Value
--skill <skill-id>	Required
--json	Authoritative payload
--prompt	Optional human-readable host briefing

The host consumes the JSON, runs the skill, returns a result. The package's nimicoding closeout admits the result under typed contract validation.

Skill Result Contracts

Each skill admits results under a typed contract.

Skill	Result contract
spec_reconstruction	.nimi/contracts/spec-reconstruction-result.yaml
doc_spec_audit	.nimi/contracts/doc-spec-audit-result.yaml
audit_sweep	.nimi/contracts/audit-sweep-result.yaml
high_risk_execution	.nimi/contracts/high-risk-execution-result.yaml

A result that does not conform to its contract fails closed during admission. There is no soft acceptance.

Reader Scenario: A New Project Runs spec_reconstruction

A team adopting Nimi Coding has a project with mixed inputs and no canonical spec yet.

1. Run nimicoding start. Bootstrap admits.
2. Project chooses host. Adapter overlay selected (e.g., Codex, Claude, oh-my-codex).
3. Run nimicoding handoff --skill spec_reconstruction --json. Package emits handoff payload.
4. Host consumes payload. Reconstructs canonical tree under admitted contract.
5. Host returns result. Package admits via nimicoding closeout.
6. Validation. Every generated rule has source basis or gap-tracking; otherwise reject.
7. Canonical tree ready. Project can now adopt the methodology for high-risk work.

The reconstruction is vendor-neutral — it works with any admitted host that satisfies the host-class capabilities.

Reader Scenario: An audit_sweep Returns A Findings Ledger

A team wants a comprehensive audit of their project under the methodology.

1. Run nimicoding handoff --skill audit_sweep --json.
2. Host runs the sweep. Reads the project under admitted read scope; emits typed findings.
3. Findings ledger frozen. Result is recorded; cannot be edited later.
4. Manager reviews. Typed findings inform the next wave-admission decisions.

The ledger is evidence. Future audits can compare against it.

Reader Scenario: A high_risk_execution Run

A team wants to run substantive AI-coding work under the methodology.

1. Manager admits a packet. Frozen with all required fields.
2. Pre-implementation audit. If authority_convergence gate fires, run the audit; record PASS.
3. Run nimicoding handoff --skill high_risk_execution --json. Hands off the packet to the host.
4. Host executes. Bound by the packet; produces output.
5. Result returned. Package validates result contract.
6. Post-implementation judgement. Independent loop reviews; record judgement.
7. Closeout. Four-closure dimensions evaluated.
8. Wave closes. Or returns to revision.

This is the full execution flow. Each step is admitted; nothing is implicit.

What Skills Do Not Do

Skill operation	Forbidden
Run AI inference inside the package	Yes — host owns runtime
Mutate project canonical truth without admission	Yes
Generate output without source basis	Yes (spec_reconstruction requires source basis or gap tracking)
Continue if host capability check fails	Yes (fail closed)

Source Basis

• nimi-coding/config/skills.yaml
• nimi-coding/config/skill-manifest.yaml
• nimi-coding/methodology/skill-runtime.yaml
• nimi-coding/methodology/skill-handoff.yaml
• nimi-coding/methodology/skill-installer-result.yaml
• nimi-coding/methodology/spec-reconstruction.yaml