Skip to content

Claude Agent Ingestion Brief

Status: Active Last reviewed: 2026-06-28 Scope: Fast startup path for Claude agents after the documentation-system consolidation.

This brief tells a Claude agent how to ingest the current docs without reading the entire repository history. It is intentionally short. Follow the linked source docs for details.

Read Order

  1. AGENTS.md - hard constraints and verification commands.
  2. Current Project State - generated facts about runtime, enforcement, launch blockers, and key commands.
  3. Current Documentation State - generated doc authority, active docs, historical docs, and tracker libraries.
  4. Project Cockpit - VitePress tracker entry point.
  5. Roadmap - active, next, gated, and held product sequence.
  6. Roadmap Implementation Plan - concrete slices, start conditions, blocked scope, acceptance proof, and verification.
  7. Execute A Roadmap Slice - prompt contract for bounded execution, closure-only work, stop rules, and deferrals.
  8. Change Routing Workflow - which docs and checks move with a given change type.
  9. Release Versioning And OTA Workflow and Release Evidence Ledger - required when a change may publish OTA/build/backend artifacts or when validating what has shipped.

Only read historical logs after the active docs say you need chronology. The old milestone material lives in CLAUDE.md Milestone Archive and is not current instruction.

Current Operating Model

  • Generated state docs are source-of-truth snapshots. Do not hand-edit docs/current-project-state.md or docs/current-doc-state.md.
  • scripts/doc-state-config.mjs owns active-doc metadata and VitePress reference grouping.
  • docs/architecture/roadmap.md owns sequence and outcomes.
  • docs/architecture/implementation-plan.md owns executable build slices.
  • docs/agent-prompts/slice-goal.md owns the copyable prompt shape for opening or resuming bounded slice work.
  • docs/workflows/change-routing.md owns what to read, update, and verify for common change types.
  • docs/workflows/release-versioning-and-ota.md owns Git/EAS classification, versioning, OTA/build evidence, and production rollout rules.
  • docs/release-evidence.md owns the durable artifact ledger: update groups, build IDs, backend deploy evidence, hosted smoke outcomes, failed attempts, and explicit no-artifact decisions.
  • The VitePress tracker loads from the same Markdown/config sources rather than a manually maintained sidebar.
  • Until R4 Launch Readiness is accepted, OTA publishes go to preview only via yarn ota:preview; production OTA commands are mechanically gated.

Default Next Work

As of 2026-06-28, the default product work is R1 closure, not open-ended R1 improvement:

  1. R1-3 Share and deep-link activation - close app-link/Kakao/preview-build proof blockers one at a time.
  2. R1-4 Club-growth hosted smoke - run or record the consolidated evidence packet only after R1-3 is accepted or explicitly deferred.

Use Roadmap Implementation Plan before touching code. If a slice's start condition is not true, do the source/evidence work first rather than jumping into implementation.

For R1, start with:

bash
yarn report:r1-release-readiness

Then choose exactly one reported blocker. Stop after that blocker is closed, proven external, or split.

If the prompt explicitly says "unattended R1 closure" or "run through R1 slices while I am away," keep the one-blocker contract per iteration but loop to the next reported blocker after verification and commit/push. Stop when R1 is accepted or the remaining blockers require owner credentials, vendor-console changes, MFA/login, device access, DNS/app-link hosting changes, production artifacts, destructive data changes, unrelated dirty work, or repeated failure on the same root cause.

Verification Baseline

For docs-system changes:

bash
yarn docs:generate
yarn docs:check
git diff --check

For code-affecting work or before claiming the repo is clean:

bash
yarn check

Remember: a green yarn check is a baseline, not launch proof. Device, hosted, vendor, legal, and credential gates remain separate when the relevant slice requires them.

Prefer targeted checks during iteration. Run the full baseline once before handoff or push unless the active slice's verification ladder explicitly needs more.

Markdown remains the source of truth. Run yarn docs:check before handoff.