Context Vigilance
Search /

Getting Started

Two commands, then your first cycle.

The kit is a plugin: the practice as a skill, seven workflow commands, templates, and one small safety check. No database, no config, no server. It works the moment it's installed.

1 Install

Claude Code

In a Claude Code session:

  1. Add the kit as a plugin source.

    /plugin marketplace add lossless-group/context-vigilance-kit

  2. Install the plugin.

    /plugin install cv@context-vigilance-kit

  3. Set up context-v/ in your repo.

    /cv:init

Straight from GitHub, no approval needed. In a terminal instead of a session, use claude plugin marketplace add … and claude plugin install ….

Any other agent

Codex, Cursor, OpenCode, Pi, and other tools that read Agent Skills. In a terminal:

  1. Download the kit.

    git clone https://github.com/lossless-group/context-vigilance-kit.git

  2. Copy its skills into your tool's skills folder (this is the common one; check your tool's docs).

    cp -R context-vigilance-kit/agent-skills/* ~/.agents/skills/

Then ask your agent for a skill by name: "use the init skill."

Anything that reads AGENTS.md

No skills support at all? Paste the AGENTS.md snippet into your project. It points the agent at the practice, and that alone gets you most of the way.

2 Run your first cycle

Each command puts the agent in a role. In Claude Code you type them; in other tools, name the skill. Start in any repo, even an empty one.

  1. /cv:init scaffold

    Creates context-v/ with its folders, gitignores extra/, and offers an AGENTS.md pointer and a changelog/.

  2. /cv:explore "…" write

    Starts a doc from a template with real IDs and dates. Each doc gets its folder's type and a line in that folder's index. One shortcut per kind: /cv:spec, /cv:plan, /cv:prompt, /cv:blueprint, /cv:remind, /cv:issue, /cv:decide, /cv:handoff. Or /cv:new <kind> for any of them.

  3. /cv:prep product manager

    Walks an idea up the ladder: exploration → spec → plan, with acceptance criteria someone can check. Writes no code.

  4. /cv:implement <plan> lead engineer

    Builds one signed-off spec or plan, verifying each step by running it, and closes its status honestly.

  5. /cv:loop <plan> VP of engineering

    Builds a bigger one with subagents, under your own process, written down once in context-v/loops/.

  6. /cv:reflect closing out

    Writes the as-built state, files issues, a changelog entry and a handoff, then commits and pushes.

  7. /cv:kickoff next session

    Reads the indexes and the latest handoff, and proposes the few docs worth loading before you start.

  8. /cv:upgrade stay current

    Pulls the latest kit, says what changed, and offers (never forces) to bring your docs up to new conventions.

3 What you'll have

your-repo/
├── AGENTS.md          # points agents here
├── changelog/         # optional ship log
└── context-v/
    ├── index.md       # table of contents (OKF)
    ├── specs/
    ├── plans/
    ├── prompts/
    ├── blueprints/
    ├── reminders/
    ├── agent-skills/
    ├── explorations/
    ├── issues/
    ├── extra/         # scratch, gitignored
    └── sitemap/

Folders like loops/, handoffs/, and decisions/ appear when a workflow first needs them. Every doc starts from a template with frontmatter filled in, including two IDs generated by command.

One check runs on its own. Before your agent writes to context-v/, the plugin makes sure the frontmatter parses and the dates and IDs are well-formed. If not, the write is blocked with a message saying exactly what to fix. Everything else is judgment, and stays that way.

Not sure why something is the way it is? The patterns explain each convention and where it came from.

4 Recommended companions

Optional, and installed from their own projects so they stay current. The kit's skills use them when they're present and never require them.

Graphify

A map of your codebase: a knowledge graph built by parsing the code, with no LLM and no database in the way, so it takes seconds. /cv:kickoff reads it before scanning folders.

  1. Install the command-line tool.

    uv tool install graphifyy

  2. Add its skill to your agent.

    graphify install

safishamsi/graphify →

Archify

Architecture, workflow, and sequence diagrams checked against your repo. /cv:prep can put one in a spec; /cv:reflect can draw the as-built version.

  1. npx skills add tt-a1i/archify -g

tt-a1i/archify →

Chroma skills

Correct guidance for semantic search, for when your context-v/ grows past what folder scans handle well.

  1. npx skills add chroma-core/agent-skills --skill "chroma-local"

chroma-core/agent-skills →
Disclosure: the kit's author is an investor in Chroma.

5 When you outgrow the defaults

Use a ticket system, a team chat, a docs site? Copy starters/config.md to context-v/config.md and say which tool fills each role: tracker, chat, docs, design, memory. Skills name roles, so they work with whatever you choose. Secrets never go in it; it names environment variables instead.

The kit is MPL-2.0. What you make with it, including everything in your context-v/, is yours.