Safe re-initialization of a project's agent knowledge — burn the rotten docs, rebuild from the code and the user's intent.
clean-slate is an agent skill
for the day your knowledge files stop helping: CLAUDE.md contradicts the
code, three generations of agents left conflicting notes, the roadmap
describes a project that no longer exists. Past that point, patching costs
more than rebuilding.
/clean-slate .
- Inventory — finds every agent-knowledge file (CLAUDE.md, AGENTS.md, cursor rules, architecture docs, roadmaps, old plans) and classifies each delete / keep / ask. Human docs (README, CHANGELOG, end-user docs) are never in scope.
- Salvage — mines the old files for load-bearing claims and tags each
one: verified (fresh
file:linecitation — survives), refuted (becomes a "common wrong assumption" warning), or unverifiable from code (deploy quirks, house rules — carried to the interview or the roadmap's open questions, never silently dropped). - Backup gate ⛔ — presents the exact delete list and asks where to back
it up (recommended:
docs/_archive/<date>/+ a git tag). Nothing is deleted without an explicit answer. Headless runs take the recommended backup automatically — and if it can't be verified, they stop instead of deleting. - Clean — backs up, count-verifies the backup against the delete list (a command, not a judgment call), deletes only the confirmed list.
- Fresh read — inventories the codebase from source: structure, wiring, data flow, git trajectory. Old docs are gone and are not trusted from memory.
- Direction interview — asks the user where the project should go: one question at a time, each with a recommended answer grounded in the code, five questions max.
- Rebuild — writes a strong, thin, fully-cited knowledge set:
AGENTS.md(under 150 lines),docs/ARCHITECTURE.md,ROADMAP.md— including one anti-rot rule aimed at whatever made the old docs rot, so the reset doesn't just reset the clock. - Report — deleted / backed up / salvaged / dropped-as-false / created.
The philosophy is shared with the deep-plan skill family: no claim without evidence — and additionally: no deletion without a confirmed backup decision. A thin true document beats a thick maybe-true one.
npx skills add silkyland/clean-slateOr copy this directory into your agent's skills folder
(e.g. ~/.claude/skills/clean-slate/).
clean-slate/
├── SKILL.md # 8-step workflow + safety rules
└── references/
├── knowledge-file-patterns.md # What's in scope, what's untouchable
├── interview-guide.md # The direction interview question bank
└── rebuild-templates.md # AGENTS.md / ARCHITECTURE.md / ROADMAP.md structures
Follows the Vercel skills single-skill layout and Anthropic's skill authoring best practices.
| Skill | Moment |
|---|---|
| know-my-repo | Day one: onboard onto a repo with zero knowledge |
| deep-plan | Plan the next feature/refactor — evidence-gated, 7 phases |
| deep-plan-ingest | Distill an accepted plan into living knowledge files |
| clean-slate | Reset rotten knowledge files — backup-gated |
| transform-my-repo | Change the architecture: migration feasibility + strategy |
| twin-my-site | Extend the web product with a native mobile twin |
| jury-my-repo | Multi-agent adversarial audit with a verified verdict |
| love-me-love-my-docs | A user manual that regenerates itself |
| seed-ah | Fake-but-production-like demo data with a manifest |
| create-my-team | Spawn and manage a subagent team for any mission |
| reproduce-my-bug | Prove the bug before anyone fixes it |
Shared law: no claim without evidence — every skill in the family is gated on citations, verification, and honest degraded modes.
MIT