desk-flip — Skill Brief

Overview

Autopsy a stuck project, extract salvageable artifacts (lessons, valid specs, non-code assets), and route the user to a fresh start in a new repo via /bootstrap-repo.

This skill exists because the old workflow — infrastructure-first, chat-driven, late testing — produces feature-complete projects that are extremely hard to test and align on between agent and user. Rather than attempting incremental migration (which preserves the structural problems), this skill always desk-flips: the old project stays in its repo as-is, and a new project begins from scratch in a separate repo using the new file-backed, prototype-first pipeline.

Goals

• Give stuck projects a clean exit with extracted value
• Produce an autopsy report that captures what went wrong and what's worth preserving
• Extract salvageable specs, designs, and non-code assets so the fresh start doesn't re-discover domain knowledge
• Route cleanly to /bootstrap-repo with extraction output as input

Non-Goals

• Incremental migration or retrofitting the new workflow onto the old codebase
• Modifying, archiving, or deleting the old project's code or repo
• Executing the fresh start — that's /bootstrap-repo's job
• Carrying forward code, dependencies, or infrastructure from the old project

Skill Contract

• Name: desk-flip
• Location: global/claude/desk-flip/SKILL.md (repo-managed, global)
• Trigger: Manual invocation only — /desk-flip or /desk-flip <project-path>
• Non-triggers: Do not invoke for greenfield projects, projects already using the new workflow, or code/dependency migrations (use /migrate for those)
• Strategy: Always desk-flip. No assessment of "should we migrate vs. restart" — if the user invoked this skill, the answer is restart.

Workflow

1. Identify the stuck project.

• Accept an optional project path argument; default to current working directory.
• Verify it's a git repo with existing code (not empty).
• Read CLAUDE.md, README.md, tasks/, specs/, docs/ if they exist.

1. Run the autopsy.

• Use subagents to parallelize analysis across:
• Lessons extraction: What went wrong? Scan git history, commit patterns, any existing tasks/lessons.md, README, and CLAUDE.md for evidence of the old workflow's failure modes (infrastructure-before-prototype, late testing, alignment drift).
• Spec/design extraction: Identify specs, designs, research docs, and interview logs that capture valid domain knowledge, user needs, or product intent — regardless of whether the implementation matched.
• Asset inventory: Find non-code assets worth preserving — icons, copy, research documents, API contracts, design tokens, brand assets.
• Do NOT extract code, dependencies, database schemas, or infrastructure config.

1. Write the autopsy report.

• Produce desk-flip-report.md with sections:
• ## Project Summary — what the project was, what state it's in
• ## What Went Wrong — specific failure modes with evidence
• ## Salvageable Specs & Designs — list with file paths and one-line descriptions of what's still valid
• ## Salvageable Assets — non-code assets with file paths
• ## Lessons for the Fresh Start — actionable guidance for the new project to avoid the same mistakes
• ## Recommended Bootstrap Input — a condensed brief suitable as input to /bootstrap-repo
• Present the report to the user for review before routing.

1. Route to fresh start.

• Recommend the user create a new repo and run /bootstrap-repo with the extraction output.
• Do NOT create the new repo or run bootstrap-repo — the user does this.

Inputs and Outputs

Inputs:

• Optional: project path (defaults to cwd)
• The existing project's codebase and git history

Outputs:

• desk-flip-report.md — autopsy report with extraction (written to the old project repo for reference)
• Terminal output: summary + recommended next command

Safety and Side Effects

• Read-only on the old project — no modifications, no archival, no branch creation. The old repo stays exactly as-is.
• The only file written is desk-flip-report.md in the old project root.
• Does not create the new repo — the user handles that.
• Does not commit or push (no meaningful mutations to ship).
• If the project doesn't look stuck (already has new-workflow artifacts and appears healthy), warn the user and confirm before proceeding.

Verification and Benchmark Coverage

Deterministic (layer 1):

• Verify desk-flip-report.md is produced with all required sections
• Verify no other files in the old repo are modified
• Verify extraction doesn't include code files or dependency lists

Subjective (layer 2):

• Agent review: Are the "What Went Wrong" findings specific and evidence-backed, not generic?
• Agent review: Are the salvageable specs correctly identified (not stale/invalid ones)?
• Agent review: Is the "Recommended Bootstrap Input" brief actionable for /bootstrap-repo?

Related Skills

Open Questions

1. Should the autopsy report format be standardized enough that /bootstrap-repo can parse it automatically, or is human copy-paste fine?
2. Should the skill maintain a registry of desk-flipped projects (old repo → new repo mapping) for future reference?
3. Should there be a lightweight "pre-flight" mode that just tells the user what it would extract without writing anything?

Assumptions & Risks

• Assumption: The user has already decided to desk-flip. This skill doesn't try to talk them out of it.
• Assumption: The old repo stays accessible for future reference (it's not deleted).
• Risk: Extraction quality depends on how well the old project documented itself. Projects with no specs/docs will produce thin reports.
• Risk: The user may expect code to carry forward. The skill explicitly does not do this — make this clear upfront.

Recommended Creation Route

/create-agentic-skill desk-flip