Kickoff

Kickoff

Project context kickoff. Conversational interview that builds structured docs for AI agents.

Agent Capabilities

Fallback rule: If codebase access is limited, the skill works from conversation alone. Warn the user and proceed.

Actions

Flow

First time (kickoff):
  init → todo (answer open questions) → generate

With existing codebase:
  init (bootstraps from artifacts) → todo → generate

Maintenance:
  update {axis} → generate

Enrich context:
  todo → generate

Health check:
  review → update (if drift detected) → generate

Strategic challenge:
  analyze → update (if insights) → generate

Quick regeneration:
  generate (reads existing .context/)

Each action is standalone — usable independently without requiring prior actions.

Philosophy

• Kickoff, not questionnaire: The interview is a conversation. Agent adapts to answers, challenges vague responses, and knows when to stop.
• Objectives over scripts: Axis templates define WHAT to know, not WHAT to ask. The agent decides HOW based on the conversation.
• Bootstrap first: Read existing artifacts (package.json, README, git log, file structure) BEFORE asking questions. Interview fills gaps, not blanks.
• Checkpoint per axis: Write docs immediately after each axis completes. Session can be interrupted and resumed.
• TBD is a valid answer: Early-stage projects can't answer everything. "Not decided yet" is captured explicitly, never fabricated.
• Intention over description: Capture WHY decisions were made, not just WHAT exists. Agents can infer the "what" from code — they can't infer the "why".
• Merge, never overwrite: CLAUDE.md generation always diffs against existing content. Never silently destroy user's custom sections.
• Adapt to project: Not every project needs 7 axes. Detect project type and propose relevant axes.

Axes

Interview Conduct

The agent follows the rules in interview.md. Key rules:

• Max 2 questions per message
• Max 3 follow-ups per objective before accepting TBD
• Product and Engineering: insist and challenge vague answers
• Other axes: accept "not defined yet" and move on
• Reformulate understanding before moving to next axis
• Generate axis docs immediately when axis is complete (checkpoint)
• Never invent information — if it wasn't said, it's not written

Bootstrap Protocol

Before starting the interview, scan the project for existing artifacts:

1. SCAN sources (in order):
   - package.json / Cargo.toml / pyproject.toml → stack, deps, scripts
   - README.md → project description, setup instructions
   - Git log (last 20 commits) → recent activity, conventions
   - File structure (top 2 levels) → architecture patterns
   - Existing CLAUDE.md → current context, conventions
   - .env.example → services, integrations
   - CI config (.github/workflows, etc.) → infrastructure

2. STORE scan results as bootstrap_context (do NOT write files yet):
   - Pre-fill data for stack, architecture, conventions
   - Note what was auto-detected vs what needs confirmation
   - Files are written only at checkpoint time (after axis interview + user confirmation)

3. INTERVIEW fills gaps:
   - Show what was auto-detected: "I found X — is this accurate?"
   - Ask only about what couldn't be inferred
   - Product axis is always fully conversational (can't be inferred from code)

Action Router

User input
  |
  |- "init", "kickoff", "kickoff init",
  |  "project setup", "onboard project"       -> Load references/actions/init.md
  |
  |- "update product", "update engineering",
  |  "kickoff update {axis}"                  -> Load references/actions/update.md
  |
  |- "todo", "kickoff todo",
  |  "open questions", "@TODO.md"             -> Load references/actions/todo.md
  |
  |- "analyze context", "challenge context",
  |  "kickoff analyze"                        -> Load references/actions/analyze.md
  |
  |- "review context", "check context",
  |  "kickoff review"                         -> Load references/actions/review.md
  |
  |- "generate", "generate claude",
     "kickoff generate"                       -> Load references/actions/generate.md

Loading rule: Read the action file BEFORE executing. The action file contains all logic, references, and templates needed.

Output Structure (in user's project)

.context/
|-- product/
|   |-- vision.md
|   |-- users.md
|   |-- features.md
|   +-- roadmap.md
|-- engineering/
|   |-- stack.md
|   |-- architecture.md
|   +-- conventions.md
|-- decisions/
|   |-- 001-{decision-slug}.md
|   +-- ...
|-- design/                   (if UI)
|   |-- principles.md
|   +-- tokens.md
|-- business/                 (if commercial)
|   |-- model.md
|   +-- landscape.md
|-- marketing/                (if PLG/content)
|   +-- positioning.md
|-- team/                     (if >1 person)
|   +-- team.md
|-- TODO.md
+-- _meta.md

Principles

Templates

Axis Templates (interview objectives + strategies)

File Templates (generated output files)

Output Templates (chat status messages)

Configuration

All behavior is configurable by editing the skill files directly.

References

Actions:

• Init | Update | Todo | Analyze | Review | Generate

Principles:

• Interview Conduct

Templates (axes):

• Product | Engineering | Decisions | Design | Business | Marketing | Team

Templates (files):

• Context File | CLAUDE.md | Meta | Decision Record

Output (chat):

• Init | Update | Analyze | Review | Generate