agentic-workflow-auto-implement

name	agentic:workflow:auto-implement
description	Use when a developer wants to go from rough idea to working code autonomously. Triggers - imperfect prompt, solo dev wants full pipeline, "just build it", idea to implementation without manual steps.
argument-hint	[<idea.md | #issue | URL | inline description>]

/agentic:workflow:auto-implement - Idea to Implementation (Autonomous)

Usage: /agentic:workflow:auto-implement [<input>]

Take an imperfect prompt and autonomously go from idea to working code. Gathers codebase context, makes product decisions, creates a technical plan, then implements — all without manual intervention.

For developers and solo devs. Not a PM workflow — product decisions are made autonomously with documented assumptions.

Arguments

• No args: prompt for idea
• path/to/idea.md: use existing notes file
• #123 or https://github.com/.../issues/123: fetch GitHub issue
• Inline text: direct description

Workflow Overview

1. Input Detection -> 2. Architect Context Gathering (subagent) -> 3. PM Autonomous Decisions (subagent) -> 4. Generate Technical Plan (subagent) -> 5. Launch Implement Workflow

digraph workflow {
    rankdir=TB;
    input [label="Input Detection" shape=box];
    context [label="Architect Gathers\nTechnical Context +\nFunctional Understanding\n(subagent)" shape=box];
    pm [label="PM Makes Product\nDecisions Autonomously\n(subagent)" shape=box];
    plan [label="Architect Generates\nTechnical Plan\n(subagent)" shape=box];
    implement [label="Launch\nImplement Workflow" shape=box];
    done [label="DONE" shape=doublecircle];

    input -> context;
    context -> pm;
    pm -> plan;
    plan -> implement;
    implement -> done;
}

---

MANDATORY DELEGATION RULE

You MUST delegate agent work using the Task tool. You MUST NOT perform agent work yourself.

When a step says "delegate to Architect" or "delegate to PM", you:

1. Use Task tool to spawn subagent
2. Pass prompt (which tells subagent to read its own instructions)
3. Wait for result
4. Validate output exists
5. Update workflow state

You NEVER:

• Explore codebase yourself (delegate to Architect)
• Make product decisions yourself (delegate to PM)
• Write technical plans yourself (delegate to Architect)
• Write code yourself (delegate to Software Engineer via implement workflow)

If you catch yourself doing agent work, STOP and use Task tool.

Subagent Invocation Pattern

Always use general-purpose subagent type:

Task(subagent_type="general-purpose", prompt="You are the Architect agent. Read .claude/agents/agentic-agent-architect.md for your full instructions. {task-specific context}")
Task(subagent_type="general-purpose", prompt="You are the PM agent. Read .claude/agents/agentic-agent-pm.md for your full instructions. {task-specific context}")

Direct agents (steps 2-4): agentic:agent:architect, agentic:agent:pm

Transitive agents (via implement workflow, step 5): agentic:agent:software-engineer, agentic:agent:test-engineer, agentic:agent:qa, agentic:agent:test-qa, agentic:agent:security-qa

---

EXECUTION STEPS

Execute each step in order by reading the corresponding step file.

Step	File	Description
1	steps/step-01-input-detection.md	Parse args, classify input, initialize state
2	steps/step-02-architect-context.md	Delegate context gathering + functional understanding to Architect
3	steps/step-03-pm-decisions.md	Delegate autonomous product decisions to PM
4	steps/step-04-technical-planning.md	Delegate technical plan generation to Architect
5	steps/step-05-launch-implement.md	Launch implement workflow with generated plan

Start by reading steps/step-01-input-detection.md and follow NEXT STEP at end of each file.

---

AUTONOMOUS-ONLY WORKFLOW

This workflow is always autonomous. No interactive mode. The entire point is to go from rough idea to code without manual intervention.

All decisions are logged in decision-log.md with confidence scores. Low-confidence decisions (<90%) are flagged for post-implementation review.

---

Decision Logging Protocol

When subagents make autonomous decisions, they append to decision-log.md:

### DEC-{number}: {Brief Title}

**Step**: {current_step_name}
**Agent**: {deciding_agent}
**Timestamp**: {ISO timestamp}

**Context**:
{What question or ambiguity arose}

**Options Considered**:
1. {Option A} - {pros/cons}
2. {Option B} - {pros/cons}

**Decision**: {Chosen option}

**Confidence**: {percentage}%

**Rationale**:
{Why this choice was made}

**Trade-offs**: {what was sacrificed}
**Reversibility**: {Easy | Medium | Hard}

---

---

TEMPLATES

Template	Purpose
templates/workflow-state.yaml	Workflow state tracking schema

---

ARTIFACTS

All outputs: .claude/_agentic_output/task/auto-implement/{topic}/{instance_id}/

Artifact	Description
workflow-state.yaml	Workflow state tracking
decision-log.md	All autonomous decisions from all agents
input-idea.md	Original idea/prompt
technical-context.md	Codebase analysis from Architect
functional-understanding.md	Plain-language behavior synthesis from Architect
product-decisions.md	Product scope, acceptance criteria from PM
technical-plan.md	Implementation plan from Architect

---

ERROR HANDLING

Context Gathering Fails

If Architect returns incomplete context:

1. Re-run with more specific guidance
2. If second attempt fails, proceed with partial context (log as LOW_CONFIDENCE)

PM Cannot Make Confident Decisions

If PM confidence < 90% on critical product decisions:

1. Log as LOW_CONFIDENCE in decision-log.md
2. Add to Open Questions section
3. Proceed with best-guess — developer reviews post-implementation

Plan Generation Fails

If Architect cannot produce valid plan:

1. Re-run with explicit guidance about what's missing
2. If second attempt fails, HALT and present partial artifacts to developer

Implement Workflow Fails

If implement workflow fails mid-execution:

1. Log error in workflow-state.yaml
2. Set status: "failed"
3. Present all artifacts generated so far for developer review

Step Failure

If any step fails:

1. Log error in workflow-state.yaml
2. Set status: "failed"
3. Attempt recovery once, then halt with detailed error log

---

Execution

Start workflow by reading step 1:

Read steps/step-01-input-detection.md

Follow each step file's instructions sequentially. Each step ends with a reference to the next step.