The execution map that keeps AI agents from losing context on long, multi-step work.
Agents rarely fail because one step is hard. They fail because they lose the thread:
- Forget a decision made twenty steps ago
- Never notice that step 12 actually depends on step 5
- Mark a phase "done" while its real completion criterion was never met
- Drift into a side detail far from the original goal
These are context management failures, not effort failures. The fix isn't "focus harder" — it's an external, written, relationship-explicit map the agent can re-read at any moment.
Spine is a portable Agent Skill that turns any complex, interdependent request into a living tree map instead of a flat to-do list. Every phase has explicit relationships, a verifiable completion criterion, and one file — SPINE_MAP.md — that is the single source of truth, updated after every completion.
| Step | Name | Output |
|---|---|---|
| 0 | Structural understanding | A relationship map of what already exists (or the conceptual domain, for new work) |
| 1 | Atomic decomposition | Small, independently verifiable tasks — one completion criterion each |
| 2 | Tree map | Numbered phases connected by four explicit relationship types |
| 3 | Phase plans | Goal, completion criterion, tasks, and execution steps for each phase |
| 4 | Master plan | One document — SPINE_MAP.md — the single source of truth |
| 5 | Execution discipline | No drift, no forgetting, no premature closure |
A flat numbered list assumes step N only depends on step N-1. Real projects don't work that way. Spine makes the graph explicit and machine-checkable:
| Type | Meaning |
|---|---|
depends_on |
This phase cannot start until the prerequisite phase is complete |
feeds |
This phase's output is consumed by a phase further downstream |
parallel_with |
Explicitly declared safe to run concurrently with another phase |
conflicts_with |
Touches the same resource as another phase — must be serialized |
Spine follows the open Agent Skills format — a folder with a SKILL.md file that any compatible agent can load. Pick whichever install path fits your setup.
openskills detects your installed agents (Claude Code, Cursor, Windsurf, Cline, Codex, and anything that reads AGENTS.md) and installs into the right directory automatically:
npx openskills install Open2Loop/spine-skillClone the repo, then copy the spine/ folder into your agent's skills directory:
git clone https://github.com/Open2Loop/spine-skill.git| Agent | Personal (global) | Project-local |
|---|---|---|
| Claude Code | cp -R spine ~/.claude/skills/ |
cp -R spine .claude/skills/ |
| Cursor | cp -R spine ~/.cursor/skills/ |
cp -R spine .cursor/skills/ |
| Codex CLI | cp -R spine ~/.codex/skills/ |
cp -R spine .agents/skills/ |
| Windsurf | cp -R spine ~/.windsurf/skills/ |
cp -R spine .windsurf/skills/ |
| Cline (VS Code) | cp -R spine ~/.cline/skills/ |
cp -R spine .cline/skills/ |
| Gemini CLI | cp -R spine ~/.gemini/skills/ |
cp -R spine .gemini/skills/ |
| OpenCode | cp -R spine ~/.opencode/skills/ |
cp -R spine .opencode/skills/ |
| Any AGENTS.md-based agent | — | cp -R spine .agents/skills/ |
Not seeing your agent? Copy
spine/into.agents/skills/in your project root — it's the vendor-neutral convention several agents fall back to, or check your agent's own skills/plugin documentation for its exact directory.
Use $spine to decompose this feature into SPINE_MAP.md, then execute phase by phase without losing context.
Use $spine to build a new user registration flow via phone number. Map existing auth patterns first, then execute.
Use $spine on this multi-service refactor. I need explicit depends_on/feeds between phases and validate_map.py passing before any phase is marked done.
Use $spine to design the user flow for checkout. Start with the structural map (step 0) before writing any code.
After building or updating SPINE_MAP.md:
python3 spine/scripts/validate_map.py SPINE_MAP.mdThe validator checks for: dependency cycles, references to phases that don't exist, incomplete phases, mismatched depends_on/feeds pairs, phases marked done while a dependency isn't, and unresolved "needs confirmation" markers.
spine/
├── SKILL.md # Core runtime instructions
├── agents/openai.yaml # Agent interface metadata
├── assets/
│ ├── spine-logo.svg
│ ├── readme-hero.svg
│ ├── six-steps.svg
│ ├── relationship-types.svg
│ ├── roadmap-template.md # Copy this to SPINE_MAP.md to start
│ └── example-user-flow-map.md # Full worked example (phone signup)
├── references/
│ └── methodology.md # Deep reference: cycles, parallel agents, recovery
└── scripts/
└── validate_map.py # Structural validator
Use Spine when:
- Building a feature, system, or product
- Designing a user flow
- Any interdependent, multi-part task — even if the request didn't say "plan" or "roadmap"
Skip Spine when:
- The task is a single, unambiguous step ("change this button's color", "fix this typo")
SKILL.md is the core runtime and stays small. Deeper references (methodology.md, templates, examples) load only when the task actually needs them — sharp context without paying for every rule on every request.
MIT — see LICENSE.
Spine · by Open2Loop · open2loop.github.io/spine-skill