name	magic-spec
description	Workflow for creating and managing project specifications.
handoffs	[{"label":"Generate tasks","workflow":"magic-task","prompt":"Proceed to task generation without interruption to maintain Zero-Prompt workflows (Rule C9).","condition":"registry_updated"},{"label":"Add a rule","workflow":"magic-rule","prompt":"Add a project-wide convention discovered during spec work.","condition":null}]

Specification Workflow

Write Permissions (Hard Limit)

Allowed write target — ONLY .design/ subtree. Everything outside .design/ is FORBIDDEN. No exceptions.

If you are about to write to any file outside .design/ — STOP. This is out of scope for magic.spec.

Triggers: "Create spec", "Update spec", "Explore", "Brainstorm", "Review registry", "Check specs", "Verify specs" Scope: Architectural exploration and specification authoring — what exists and how it is structured. Hints:

Explore Mode: Used for safe brainstorming. Transition to writing AUTOMATICALLY upon specific input or if Anti-Stall triggers (≥1 question asked without file creation). Do not stall.
Delta Edits: Use surgical search-and-replace tools for specs >200 lines to prevent corruption.
T4 Rule Capture: When input contains "remember that..." / "project rule:", spec workflow applies Tier Routing (global vs workspace RULES.md) and Duplication Check before writing — see §T4 Inline Guards in full implementation.
Pipeline: magic.spec → magic.task → magic.run (orchestration by magic.task, execution by magic.run)
Finalization: After dispatch, run node .magic/scripts/executor.js finalize --workflow=spec and display output verbatim. Never auto-commit. See .magic/spec.md §Finalization Protocol.

Full implementation: .magic/spec.md · Skill: skills/magic-spec/SKILL.md Read that file before proceeding. Do not execute any steps until it is read. Executor: Use node .magic/scripts/executor.js <script> for all automation. Anti-Hallucination Guard: Do not invent or execute any ad-hoc physical scripts (e.g., custom .js, .sh test runners) for internal engine operations. Magic SDD workflow steps are intended to be evaluated cognitively by the LLM unless an executor script is explicitly provided.

Specification Workflow

Write Permissions (Hard Limit)

Allowed write target — ONLY .design/ subtree. Everything outside .design/ is FORBIDDEN. No exceptions.

If you are about to write to any file outside .design/ — STOP. This is out of scope for magic.spec.

Explore Mode: Used for safe brainstorming. Transition to writing AUTOMATICALLY upon specific input or if Anti-Stall triggers (≥1 question asked without file creation). Do not stall.

Delta Edits: Use surgical search-and-replace tools for specs >200 lines to prevent corruption.

T4 Rule Capture: When input contains "remember that..." / "project rule:", spec workflow applies Tier Routing (global vs workspace RULES.md) and Duplication Check before writing — see §T4 Inline Guards in full implementation.

Pipeline: magic.spec → magic.task → magic.run (orchestration by magic.task, execution by magic.run)

Finalization: After dispatch, run node .magic/scripts/executor.js finalize --workflow=spec and display output verbatim. Never auto-commit. See .magic/spec.md §Finalization Protocol.

Full implementation: .magic/spec.md · Skill: skills/magic-spec/SKILL.md Read that file before proceeding. Do not execute any steps until it is read. Executor: Use node .magic/scripts/executor.js <script> for all automation. Anti-Hallucination Guard: Do not invent or execute any ad-hoc physical scripts (e.g., custom .js, .sh test runners) for internal engine operations. Magic SDD workflow steps are intended to be evaluated cognitively by the LLM unless an executor script is explicitly provided.