arc-using is a bounded router for ArcForge skills. It helps choose the smallest useful workflow for the current task. It is guidance, not a global law.

Use it when:

The user asks to use ArcForge or an ArcForge skill.
The task is an ArcForge workflow task: brainstorming, refining specs, planning, implementing epics, verifying, evaluating, or maintaining ArcForge skills.
You are unsure which ArcForge skill should handle the next step.

Respect higher-priority instructions, explicit user constraints, and the host harness. If a workflow would add more friction than value, do not force it.

How to Access Skills

In Claude Code: Use the Skill tool. When you invoke a skill, its content is loaded and presented to you — follow it directly. Never use the Read tool on skill files.

In other environments: Use the platform's skill-loading mechanism, or read the relevant skill documentation when no tool exists.

Core Philosophy

Skills are tools, not laws.
Prefer the smallest useful workflow.
File artifacts are the source of truth when the workflow creates them.
Session context is current workflow state only; resume from files, not memory.
Strong workflows are opt-in by task fit, not always-on behavior.

File Artifacts = Truth

SDD pipeline v2 uses per-spec layout:

docs/plans/<spec-id>/<YYYY-MM-DD>/design.md → Design documents
specs/<spec-id>/spec.xml + specs/<spec-id>/details/*.xml → Refined specifications
specs/<spec-id>/dag.yaml + specs/<spec-id>/epics/ → Implementation plans
Isolated feature work → see Worktree Rule below

Worktree Rule

ArcForge worktrees live at ~/.arcforge/worktrees/<project>-<hash>-<slug>/, computed at runtime by ${ARCFORGE_ROOT}/scripts/lib/worktree-paths.js. Two tiers exist: epic worktrees carry an .arcforge-epic marker and are coordinator-managed; generic worktrees (experiments, hotfixes, review checkouts) carry no marker.

When touching worktrees:

Don't hardcode paths in output. Use abstract language like "the worktree" or fill paths from CLI output.
Don't create worktrees manually. For epic work, delegate to arc-coordinating expand; for everything else, delegate to arc-using-worktrees — so marker schema and DAG sync stay valid.
Re-entering an epic worktree: read its absolute .path from arcforge status --json (the epic-tier surface); do not reconstruct paths from memory.
Locating a generic worktree: read its path from arcforge worktree list --json (the generic surface) — never status --json, which only knows epic worktrees.
Direct file-editing belongs in a worktree; base sessions coordinate, worktree sessions implement. The .arcforge-epic marker distinguishes an epic worktree from a generic one.

For derivation rules, marker schema, and cleanup semantics, see docs/guide/worktree-workflow.md.

Routing Flow

digraph skill_flow {
    "User message received" [shape=doublecircle];
    "ArcForge workflow task?" [shape=diamond];
    "Simple/read-only/eval/grading?" [shape=diamond];
    "Choose smallest useful skill" [shape=box];
    "Invoke/read skill" [shape=box];
    "Proceed directly" [shape=box];
    "Act with evidence" [shape=doublecircle];

    "User message received" -> "ArcForge workflow task?";
    "ArcForge workflow task?" -> "Proceed directly" [label="no"];
    "ArcForge workflow task?" -> "Simple/read-only/eval/grading?" [label="yes"];
    "Simple/read-only/eval/grading?" -> "Proceed directly" [label="yes"];
    "Simple/read-only/eval/grading?" -> "Choose smallest useful skill" [label="no"];
    "Choose smallest useful skill" -> "Invoke/read skill";
    "Invoke/read skill" -> "Act with evidence";
    "Proceed directly" -> "Act with evidence";
}

Skill Priority

When multiple skills could apply, choose the smallest useful one:

Clarify intent — arc-brainstorming when requirements or decisions are unclear.
Formalize source of truth — arc-refining when converting a design/decision log into structured specs.
Plan work — arc-planning when a refined spec needs an implementation DAG.
Execute work — arc-coordinating, arc-dispatching-teammates, arc-looping, or arc-implementing based on DAG/worktree context.
Cross-cutting quality — use discipline skills only when their trigger is actually present.

Examples:

"Let's build X" → arc-brainstorming if design is unclear; arc-planning if a refined spec already exists.
"Fix this bug" → arc-debugging if cause is unknown; arc-tdd if cause and expected behavior are clear.
"Implement epic" → arc-planning if no specs/<spec-id>/dag.yaml; coordination/implementation skills if the DAG exists.
"Audit this skill/workflow" → arc-evaluating when shipping/merge/completion evidence matters.

Execution & Finishing Choosers

When two skills cover the same step, pick by the concrete condition:

Decision	Pick
Run a prepared task list	`arc-executing-tasks` (human checkpoints per batch) vs `arc-agent-driven` (fresh subagent per task + two-stage review)
Dispatch parallel work	`arc-dispatching-parallel` (independent features, one worktree) vs `arc-dispatching-teammates` (multi-epic via DAG, lead present)
Set up an isolated workspace	`arc-coordinating expand` for epic work (DAG-tracked, marker'd); `arc-using-worktrees` for a generic worktree (experiment, hotfix, review checkout — any repo, no DAG)
Finish work	`arc-finishing` (Step 0 discriminates on `.arcforge-epic`: epic path = coordinator merge; non-epic path = 4-option gate)

Full skill catalog: README "What's Inside" or docs/guide/skills-reference.md.

Discipline Skills — Conditional Triggers

These skills activate during a workflow when the condition is present. They are not mandatory pipeline steps for every message.

Condition	Skill	Gate
About to write implementation code	`arc-tdd`	Failing test before production code
Test fails or unexpected behavior appears	`arc-debugging`	Root cause before fixes
About to claim work is complete	`arc-verifying`	Fresh verification evidence before completion claims
Task or feature complete	`arc-requesting-review`	Review before proceeding to next task
Received code review feedback	`arc-receiving-review`	Technical rigor, not performative agreement
User asks about vault health, missing links, or orphan notes	`arc-maintaining-obsidian` audit mode	Propose changes, never auto-modify without approval
About to ship, merge, or mark complete a skill, agent, or workflow	`arc-evaluating`	Eval evidence that does not return `INSUFFICIENT_DATA`

Instinct & Learning Routes

Five skills touch the diary/instinct system. Route by the concrete trigger, not by the word "remember" — they are distinct entry points, not interchangeable:

User intent	Skill
Capture THIS session's reflections as a diary entry	`arc-journaling`
Extract recurring patterns from 5+ accumulated diaries	`arc-reflecting`
Manually save ONE insight as an instinct right now	`arc-recalling`
Review / confirm / contradict auto-detected instincts	`arc-observing`
Review the learning-candidate queue (when optional learning is enabled)	`arc-learning`

When Not to Route

Do not force an ArcForge workflow when the task is:

A simple factual answer or direct clarification.
Read-only inspection where no workflow decision is needed.
Harness, eval, or grading execution that must preserve isolation.
A single-skill eval where arc-using would contaminate the behavior under test.
Explicitly constrained by the user to avoid workflow overhead.
Outside ArcForge's domain.

In those cases, proceed directly, and only mention ArcForge skills if they materially help.

User Instructions

User instructions say what outcome matters. ArcForge skills can help decide how to get there, but they do not override user intent, harness constraints, or higher-priority system instructions.