Plan-Validate-Execute Pipeline

Overview

Plan-Validate-Execute is a production workflow in three commands that extends native Plan Mode: /plan-start, /plan-validate, /plan-execute. Each run feeds an ADR learning loop that reduces human interruptions over time. The guide describes it as Tier 2 production practice; install its custom commands before use.

The 3 commands

/plan-start    # PRD + dynamic research team + ADRs
/plan-validate # Structural checks + triggered specialist agents
/plan-execute  # Worktree + TDD + parallel execution + merged PR

Run /clear between each command: every command reads its inputs from disk (the plan file, ADRs), nothing needs to stay in context memory.

/plan-start: 4-phase planning

If 3+ confirmed ADRs already cover the decision, phase 2 auto-resolves with no agent.

/plan-validate: 2-layer validation

Layer 1 (structural, instant): plan format, task ordering, ADR consistency. Layer 2 (specialists, 0 to 8 agents by risk): security-reviewer, db-migration-reviewer, performance-reviewer, among others, each triggered by a precise criterion (auth, schema migration, new route…).

Every issue must be resolved: auto-resolved from an existing ADR or PATTERNS.md entry, or settled by a human, which becomes a rule for next time.

/plan-execute: to a merged PR

Isolated worktree, TDD scaffolding, parallel execution of independent tasks (commit per task), quality gate (tests + smoke test), PR and squash merge, worktree cleanup. If the quality gate fails: up to 3 auto-fix attempts by dedicated debug agents, then human notification.

ADR learning loop

Every significant architectural decision generates an ADR. 1 ADR: watching. 2 ADRs: recommended default. 3+ ADRs: confirmed, auto-resolved on future plans, candidate for promotion to a hard rule in CLAUDE.md. The guide claims roughly 80% auto-resolution on a 20-plan project; no reproducible benchmark is supplied.

When to use this pipeline

Cost

Measure all planning, validation, implementation and retry calls. Current standard rates per million input/output tokens are $2/$10 for Sonnet 5.5 and $4/$20 for Opus 5.5; a task label alone does not determine the pipeline cost.

Auto plan mode

An --append-system-prompt request does not enforce native Plan Mode. Use permissions to restrict tools:

alias claude-safe='claude --append-system-prompt \
  "$(cat ~/.claude/auto-plan-mode.txt)"'

The auto-plan-mode.txt file instructs Claude to present a plan and wait for explicit approval before each tool. Useful for sessions on critical or unfamiliar codebases, outside the 3-command pipeline.