Kickoff

Kickoff — Full Workflow Orchestrator

Single command to run the complete Research → Analyze → Plan → Code → Review → Learn → Compound pipeline.

When to Use

• Starting a non-trivial feature or project
• Facing a hard problem where existing solutions might exist
• Want the full power of the research + compound workflow
• Any time you'd normally jump straight to coding on something complex

Invocation

/kickoff [problem or feature description]

Examples:
/kickoff add real-time collaborative editing to the app
/kickoff build a PDF invoice generator with custom templates
/kickoff implement OAuth2 with Google and GitHub providers
/kickoff fix the performance bottleneck in the dashboard API

The Pipeline

CLARITY → RECON → TEARDOWN? → ANALYZE → PLAN → EXECUTE → LEARN → COMPOUND
 ambiguity  research  study refs  prioritize  write plan  TDD build  log findings  next cycle?
  gate

Time budget: Target 60–90 min total. If any phase exceeds 20 min, checkpoint with user.

Abort: User can say "stop" or "skip to [phase]" at any gate. Respect immediately.

Phase 0: CLARITY GATE (ambiguity scoring)

Before any execution, score how well-defined the request is. This prevents the #1 waste: executing on vague requirements.

Skip if: Request already has specific file paths, clear acceptance criteria, and obvious scope (score would be < 0.20).

Ambiguity dimensions (scored 0.0 = unknown → 1.0 = crystal clear):

| Dimension | Weight | What to assess | |-----------|--------|----------------| | Intent | 30% | WHY the user wants this | | Outcome | 25% | WHAT end state they want | | Scope | 20% | HOW FAR the change should go | | Constraints | 15% | Technical or business limits | | Success Criteria | 10% | HOW completion will be judged |

Formula: ambiguity = 1 - (intent × 0.30 + outcome × 0.25 + scope × 0.20 + constraints × 0.15 + success × 0.10)

Threshold: <= 0.20 before proceeding to Phase 1.

Rules:

• Ask ONE question per round, targeting the weakest dimension
• Max 8 rounds (then proceed with explicit risk warning)
• Gather codebase facts via Glob/Grep/Read BEFORE asking user about them
• Show score after each answer so user sees progress
• Intent and scope questions come FIRST, implementation detail LAST

Example round:

Round 3 | Target: Scope | Ambiguity: 42%

I see the project has 5 document types but only field_report has the full
editing pipeline. Should this change cover all 5 types, or just field reports?

Pre-context snapshot: Save brief context to ~/.claude/artifacts/context/{slug}-{timestamp}.md before starting. This enables resuming kickoff across sessions.

Gate: When ambiguity <= 0.20, announce score breakdown and proceed to Phase 1.

Phase 1: RECON (automatic)

Dispatch the deep-recon research swarm:

Invoke: /recon [problem description]

4 parallel agents search for:
- Existing packages/solutions
- Community discussions (Reddit, HN, SO, Substack)
- Documentation and tutorials
- Source code analysis of top solutions

Output: Research brief saved to ~/.claude/artifacts/research/

Gate: Present research findings to user. Ask:

• Should we use an existing solution?
• Should we study any of these solutions deeper? (→ teardown)
• Should we build from scratch with patterns from research?

If recon finds nothing useful: Skip teardown, proceed to Analyze with what you know from the codebase.

Phase 2: TEARDOWN (conditional)

Only if research found apps/packages worth studying deeper:

Invoke: /teardown [solution found in recon]

Methods available:
- macOS app bundle analysis
- NPM/PyPI package source code analysis
- GitHub repo architecture analysis
- Web app browser analysis

Output: Teardown report saved to ~/.claude/artifacts/research/teardowns/

If teardown fails (private repo, binary-only, paywalled): Note what was inaccessible and proceed to Analyze with available information.

Gate: Present teardown findings. Confirm approach before planning.

Phase 3: ANALYZE

Score the problem and select priority:

Use the compound-engineering skill's analyze phase.

Inputs:
- Research brief from Phase 1
- Teardown findings from Phase 2 (if any)
- Current project state (errors, tests, feedback)
- Previous LEARNINGS.md entries

Output: Analysis with priority matrix and selected approach.

Phase 4: PLAN

Write implementation plan with research context:

Invoke: writing-plans skill

Inputs:
- Analysis output
- Research brief (patterns to adopt)
- Teardown findings (architecture to reference)

Every task gets:
- Specific success criteria
- Verification commands
- Reference to research findings

Output: Implementation plan with goal-driven success criteria per task.

Gate: Present plan to user for approval before execution.

Phase 5: EXECUTE

Build it using goal-driven execution:

Invoke: executing-plans or subagent-driven-development

For each task:
1. Define success criteria
2. Write test first (TDD)
3. Implement
4. Loop until ALL criteria pass
5. Commit

Quality hooks fire automatically:

• Type check after every edit
• No any types enforcement
• Auto-format
• Quality gate blocks commits until clean

Phase 6: LEARN

Update persistent learning log:

Use compound-engineering's learn phase.

Append to project LEARNINGS.md:
- What worked
- What didn't
- Patterns discovered
- Time spent per phase

Phase 7: COMPOUND

Generate cycle report and suggest next improvement:

Use compound-engineering's report phase.

Output:
- What improved (with verification evidence)
- Next priority recommendation
- Cycle metrics

If more work to do: Loop back to Phase 3 (Analyze) with user approval.

Adaptive Shortcuts

Not every problem needs every phase. The kickoff auto-adapts:

| Problem Type | Phases Used | |-------------|------------| | Complex new feature | All 7 phases | | Known problem, unclear solution | Recon → Plan → Execute → Learn | | Improvement to existing feature | Analyze → Plan → Execute → Learn | | Bug fix with research needed | Recon → Execute → Learn | | Simple bug fix | Skip kickoff, just fix it |

Detection criteria for skipping phases:

• Skip recon: You already know the solution space (internal bug fix, well-understood library)
• Skip teardown: Recon found no comparable implementations worth studying
• Skip entirely: Problem touches ≤3 files, solution is obvious, no architectural decision needed — just fix it directly

User Checkpoints

The pipeline pauses for user input at these gates:

1. After RECON — "Build, buy, or study deeper?"
2. After PLAN — "Does this plan look right?"
3. After EXECUTE — "Everything working?"
4. After COMPOUND — "Continue to next improvement?"

Model Routing

| Phase | Model | Why | |-------|-------|-----| | Recon lead | opus | Synthesizing research | | Recon agents | sonnet | Fast parallel search | | Teardown scanning | sonnet | Pattern matching | | Teardown synthesis | opus | Architecture analysis | | Analyze | opus | Priority judgment | | Plan | opus | Architectural decisions | | Execute (implementers) | sonnet | Well-defined tasks | | Learn | sonnet | Documentation | | Compound report | opus | Connecting metrics |

Integration

This skill orchestrates all other skills:

kickoff
├── deep-recon (Phase 1)
│   ├── package-research
│   └── dispatching-parallel-agents
├── app-teardown (Phase 2)
├── compound-engineering (Phases 3, 6, 7)
│   ├── writing-plans (Phase 4)
│   └── executing-plans / subagent-driven-development (Phase 5)
│       ├── test-driven-development
│       ├── verification-before-completion
│       └── goal-driven execution
└── finishing-development-branch (when done)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Lucface
• Source: Lucface/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.