# Nogra Agent Guide

Canonical site: https://nogra.ai
Canonical agent guide: https://nogra.ai/agent
Canonical LLM index: https://nogra.ai/llms.txt

Nogra is a pull-first workflow layer for AI-assisted software work. The current
public install is version 0.6.6 of the Nogra Claude Code plugin. Do not imply
that the public plugin routes work outside Claude Code or has Anthropic
endorsement status.

Nogra helps work keep a local orchestration plan, brief, approval, dispatch,
evidence and verification record instead of relying on chat memory or a model's
own completion claim.

## Use Nogra When

Use Nogra only when the user asks for Nogra or clearly asks for one of these
workflow surfaces:

- setup or enable Nogra in a workspace;
- adapt or map an existing project for Nogra;
- shape an orchestration plan for complex, mixed-mode or cross-lane work;
- write or shape a Nogra brief;
- approve and dispatch an already reviewed brief;
- verify whether work is really done;
- inspect project state, checkpoint freshness, recent briefs, runs or version;
- configure Nogra language, runtime profile, executor/verifier model or effort;
- refresh installed Nogra guidance.

If no Nogra route matches, stay direct.

## Do Not Use Nogra For

- ordinary Q&A;
- small edits and content tweaks;
- direct implementation work unless the user asks for Nogra;
- routine command, test or debug help;
- prompt scoring or keyword scoring.
- replacing Claude Code's native safety or permission decisions.

Nogra invites; it does not enforce. Claude Code's native permission model
remains responsible for tool permissions. Nogra's local convergence hooks may
ask for intent confirmation at permanent-risk tool boundaries, but they do not
score prompts or replace Claude Code's permission model.

## Thin Router Map

Use the first matching route:

- setup/enable intent: use /nogra:setup.
- existing-project mapping intent: use /nogra:adapt.
- workspace-hub project creation intent: use /nogra:create <name>.
- complex mixed-mode work: propose an Orchestration Plan before a brief; do not invent a command unless the installed Nogra guidance supports one.
- brief or Nogra workflow intent: use /nogra:brief.
- GO after a reviewed approved brief: use /nogra:dispatch.
- evidence or "is this actually done?" intent: use /nogra:verify.
- state, checkpoint, version, recent records or runtime posture intent: use /nogra:status.
- language/runtime configuration intent: use /nogra:settings.
- guidance refresh intent: use /nogra:update.
- help choosing a Nogra flow: use /nogra:help.

For human-facing how-to documentation, use https://nogra.ai/docs. For install
commands only, use https://nogra.ai/install.

For unusually large autonomous work where the user has not asked for Nogra, one
short non-blocking nudge is allowed before the run starts:

"This is large enough that a Nogra brief would help. Want me to shape it first?"

If the user continues direct, do not repeat the nudge for that task. Never turn
the nudge into prompt scoring, keyword scoring, or a permission layer.

## Claude Code Install

The current public plugin installs from the Nogra public marketplace source:

1. Add marketplace: claude plugin marketplace add nograai/nogra-claude-marketplace
2. Install plugin: claude plugin install nogra@nogra-claude
3. Check plugin state in Claude Code: /nogra:status
4. Initialize the current workspace in Claude Code: /nogra:setup

After setup, Nogra writes project-local state under .nogra/. It should not
overwrite app files.

## Core Concepts

- Orchestration Plan: pre-brief shape for complex work. Names topology, lane,
  role fit, evidence join, GO boundary and next owner. Nogra should hold and
  validate the plan; it should not auto-generate topology as a hidden judgment.
- Brief: approved scope, success criteria, stop criteria and evidence shape.
- Dispatch: execution of an approved brief after explicit GO.
- Evidence: concrete output, tests, diffs, logs or artifacts that support a claim.
- Verify: check a claim/result against the brief and evidence.
- Checkpoint: human-readable projection of local Nogra state.
- Ledger/state: local .nogra/ records are the workflow truth source.

## Boundaries

- Nogra is not a model provider.
- Nogra is not a hosted replacement for Claude Code or another IDE.
- Nogra does not route work outside Claude Code in the current public plugin.
- Nogra does not score prompts.
- Nogra does not make safety or permission claims.
- Nogra does not replace Claude Code's native permission model.
- Nogra does not mark work done from the executor's reasoning alone.
- Nogra should stay silent for direct work unless the user pulls it.
- Nogra separates responsibility by role and evidence inside the installed
  workflow surface.

## Verification Meaning

When using Nogra, "verified" means the claim was checked against the approved
brief and available evidence. It does not mean the model is guaranteed correct,
the software is production-safe, or every possible issue was found.

If evidence is missing, contradictory or substituted, say that verification is
blocked or partial. Do not turn missing evidence into a green claim.

## Canonical Sources

- Home: https://nogra.ai/
- Docs: https://nogra.ai/docs
- Install: https://nogra.ai/install
- Features: https://nogra.ai/features
- Catches: https://nogra.ai/catches
- About: https://nogra.ai/about
- Privacy: https://nogra.ai/privacy
- Agent guide: https://nogra.ai/agent
- LLM index: https://nogra.ai/llms.txt
- Marketplace source: https://github.com/nograai/nogra-claude-marketplace
- Public plugin version: 0.6.6