Epic: Project-specific harness toolkit for large-codebase agent work

[sungjunlee]

Context

Anthropic's large-codebase Claude Code guidance frames successful agent adoption as a project-specific harness problem: lean root context, local conventions, on-demand skills, hooks/scripts, plugins, MCPs, LSP, and subagents where they actually earn their keep.

This likely belongs near CraftKit because the job is not dev-backlog-specific. It is about helping a project design its own harness: what belongs in CLAUDE.md / AGENTS.md, what belongs in repo-local skills, what should become hooks/scripts, and when the package deserves plugin shape.

Initial craft-skill-spec judgment

• Artifact class: start as a skill suite design, not a new plugin immediately.
• Why not smaller: one single skill would overload context mapping, skill authoring, hook/script selection, and reassess cadence into one muddy trigger.
• Escalation note: promote to plugin only if the toolkit needs bundled hooks, MCP config, installable repo templates, or subagents as first-class surfaces.

Artifact thesis

• Job: help a project design and maintain a project-specific agent harness for large-codebase work.
• Non-goals: replacing dev-backlog, relay, or CraftKit's existing skill-design workflows; auto-editing project specs without a human gate.
• Target users/agents: maintainers building agent workflows for real codebases, especially multi-skill or multi-repo systems.
• Success condition: the output tells a maintainer exactly which artifacts to add, which to avoid, and how to evaluate whether the harness improves agent work.

Scope Candidates

• Context map guidance: what belongs in root CLAUDE.md / AGENTS.md vs subdirectory files vs skills.
• Project-local skill authoring: when a convention deserves a skill instead of more root prose.
• Hook/script guidance: deterministic checks, advisory reflections, and no silent self-editing.
• Periodic harness reassess: model/tool release and 3-6 month review cadence.
• Eval plan: prompts that prove the harness helps navigation, planning, review, and handoff.

Non-goals

• No live-source browsing on every run.
• No provider-specific core unless isolated in references.
• No new repo until the skill suite vs plugin boundary is proven.
• No automatic project rewrites.

Acceptance Criteria

• Child issues cover radar/source update, artifact-class spec, and MVP file/package shape.
• The design applies CraftKit radar: concise spine, progressive disclosure, explicit evals, calibrated autonomy.
• The first implementation path starts smaller than a plugin unless hooks/MCP/subagents are required.

[coderabbitai]

🔗 Related PRs

#3 - Add skill-radar curation layer + craft-skill-spec meta-skill [merged]
#5 - craft-handoff: pair per-session prompt with per-project rich doc [merged]
#22 - [codex] implement CraftKit review backlog [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!