Apex Spec Workflow
Scannednpx machina-cli add skill moshehbenavraham/apex-spec-system/spec-workflow --openclawApex Spec Workflow
A specification-driven workflow system for AI-assisted development that breaks large projects into manageable 2-4 hour sessions with 12-25 tasks each.
Core Philosophy
1 session = 1 spec = 2-4 hours (12-25 tasks)
Break large projects into manageable, well-scoped implementation sessions that fit within AI context windows and human attention spans.
A collection of sessions is a phase. A collection of phases is a mature/late technical PRD.
The 13-Command Workflow
The workflow has 3 distinct stages:
Stage 1: INITIALIZATION (One-Time Setup)
/initspec -> Set up spec system in project
|
v
/createprd -> Generate PRD from requirements doc (optional)
OR OR
[User Action] -> Manually populate PRD with requirements
|
v
/createuxprd -> Generate UX PRD from design docs (optional)
OR OR
[User Action] -> Manually populate UX PRD with requirements
|
v
/phasebuild -> Create first phase structure (session stubs)
Stage 2: SESSIONS WORKFLOW (Repeat Until Phase Complete)
/plansession -> Analyze project, create spec + task checklist
|
v
/implement -> AI-led task-by-task implementation
|
v
/validate -> Verify session completeness
|
v
/updateprd -> Sync PRD, mark session complete
|
+-------------> Loop back to /plansession
until ALL phase sessions complete
Stage 3: PHASE TRANSITION (After All Previous Phase's Sessions Are Complete)
/audit -> Local dev tooling (formatter, linter, types, tests, observability, hooks)
|
v
/pipeline -> CI/CD workflows (quality, build, security, integration, ops)
|
v
/infra -> Production infrastructure (health, security, backup, deploy)
|
v
/carryforward -> Capture lessons learned (optional but recommended)
|
v
/documents -> Audit and update documentation
|
v
[User Action] -> Manual testing and LLM audit (HIGHLY recommended)
|
v
/phasebuild -> Create next phase structure
|
v
-> Return to Stage 2 for new phase
Directory Structure
Projects using this system follow this layout:
project/
|-- .spec_system/ # All spec system files
| |-- state.json # Project state tracking
| |-- CONSIDERATIONS.md # Institutional memory (lessons learned)
| |-- SECURITY-COMPLIANCE.md # Security posture & GDPR compliance
| |-- CONVENTIONS.md # Project coding standards and conventions
| |-- PRD/ # Product requirements
| | |-- PRD.md # Master PRD
| | \-- phase_NN/ # Phase definitions
| |-- specs/ # Implementation specs
| | \-- phaseNN-sessionNN-name/
| | |-- spec.md
| | |-- tasks.md
| | |-- implementation-notes.md
| | |-- security-compliance.md
| | \-- validation.md
| |-- scripts/ # Bash automation (if copied locally)
| \-- archive/ # Completed work
\-- (project source files)
Monorepo Directory Structure
Monorepo projects use the same single .spec_system/ at the repo root. Sessions reference their target package in metadata (spec.md header and state.json), not in directory names.
monorepo-project/
|-- .spec_system/ # Single spec system at repo root
| |-- state.json # Includes monorepo flag + packages array
| |-- CONVENTIONS.md # Includes Workspace Structure table
| |-- CONSIDERATIONS.md
| |-- SECURITY-COMPLIANCE.md
| |-- PRD/
| | |-- PRD.md # Includes Package Map section
| | \-- phase_00/
| | |-- session_01_project_setup.md
| | |-- session_02_web_scaffold.md # Package: apps/web
| | \-- session_03_api_models.md # Package: apps/api
| |-- specs/
| | |-- phase00-session01-project-setup/ # Cross-cutting (no package)
| | |-- phase00-session02-web-scaffold/ # Scoped to apps/web
| | \-- phase00-session03-api-models/ # Scoped to apps/api
| \-- archive/
|-- apps/
| |-- web/ # Frontend package
| \-- api/ # Backend package
\-- packages/
\-- shared/ # Shared library
Key differences from single-repo:
- Session stubs gain optional
Package:annotation - spec.md headers include
Package:andPackage Stack:fields - CONVENTIONS.md includes a Workspace Structure table
- Session numbering is global within a phase (interleaved across packages)
Session Naming Convention
Format: phaseNN-sessionNN-name
phaseNN: 2-digit phase number (phase00, phase01)sessionNN: 2-digit session number (session01, session02)name: lowercase-hyphenated description
Examples:
phase00-session01-project-setupphase01-session03-user-authenticationphase02-session08b-refinements
Session Scope Rules
Hard Limits (Reject if Exceeded)
| Limit | Value |
|---|---|
| Maximum tasks | 25 |
| Maximum duration | 4 hours |
| Objectives | Single clear objective |
Ideal Targets
| Target | Value |
|---|---|
| Task count | 12-25 (sweet spot: 20) |
| Duration | 2-3 hours |
| Focus | Stable/late MVP |
Task Design
Task Format
- [ ] TNNN [SNNMM] [P] Action verb + what + where (`path/to/file`)
Components:
TNNN: Sequential task ID (T001, T002, ...)[SNNMM]: Session reference (S0103 = Phase 01, Session 03)[P]: Optional parallelization marker- Description: Action verb + clear description
- Path: File being created/modified
Task Categories
- Setup (2-4 tasks): Environment, directories, config
- Foundation (4-8 tasks): Core types, interfaces, base classes
- Implementation (8-15 tasks): Main feature logic
- Testing (3-5 tasks): Tests, validation, verification
Parallelization
Mark tasks [P] when they:
- Create independent files
- Don't depend on each other's output
- Can be done in any order
Critical Requirements
ASCII Encoding (Non-Negotiable)
All files must use ASCII-only characters (0-127):
- NO Unicode characters
- NO emoji
- NO smart quotes - use straight quotes (" ')
- NO em-dashes - use hyphens (-)
- Unix LF line endings only (no CRLF)
Validate with:
file filename.txt # Should show: ASCII text
grep -P '[^\x00-\x7F]' filename.txt # Should return nothing
Over-Arching Rules
- Complete one session at a time before starting next
- Update task checkboxes immediately as work progresses
- Follow workflow sequence - resist scope creep
- Read spec.md and tasks.md before implementing
State Tracking
The .spec_system/state.json file tracks project progress:
{
"version": "2.0",
"project_name": "Project Name",
"current_phase": 0,
"current_session": null,
"phases": {
"0": {
"name": "Foundation",
"status": "in_progress",
"session_count": 5
}
},
"completed_sessions": [],
"next_session_history": []
}
Monorepo State Tracking
When monorepo is confirmed true, state.json gains a packages array and uses object-form completed_sessions:
{
"version": "2.0",
"project_name": "Acme Platform",
"monorepo": true,
"packages": [
{ "name": "web", "path": "apps/web", "type": "frontend", "stack": "TypeScript + React" },
{ "name": "api", "path": "apps/api", "type": "backend", "stack": "Python 3.12 + FastAPI" },
{ "name": "shared", "path": "packages/shared", "type": "library", "stack": "TypeScript" }
],
"current_phase": 0,
"current_session": "phase00-session04-api-models",
"phases": {
"0": { "name": "Foundation", "status": "in_progress", "session_count": 6 }
},
"completed_sessions": [
{ "id": "phase00-session01-project-setup", "package": null },
{ "id": "phase00-session02-web-scaffold", "package": "apps/web" },
{ "id": "phase00-session03-shared-types", "package": "packages/shared" }
],
"next_session_history": []
}
The monorepo field uses three states:
| Value | Meaning | Behavior |
|---|---|---|
null (or absent) | Unknown / not yet determined | Commands look for signals, prompt if found |
true | Confirmed monorepo | Commands use package-aware logic |
false | Confirmed single-repo | Commands skip package logic entirely |
Single-repo projects see no packages field and keep string-form completed_sessions. Existing state files without monorepo are treated as null (unknown), which behaves identically to classic single-repo until the user confirms otherwise.
Command Quick Reference
| Command | Purpose | Input | Output |
|---|---|---|---|
/initspec | Initialize spec system | Project info | .spec_system/ structure |
/createprd | Generate master PRD | Requirements doc or user text | PRD/PRD.md |
/createuxprd | Generate UX PRD | Design docs or user text | PRD/PRD_UX.md |
/plansession | Analyze, spec, and task list | state.json, PRD | specs/.../spec.md + tasks.md |
/implement | Code implementation | spec.md, tasks.md | implementation-notes.md |
/validate | Verify completeness | All session files | security-compliance.md, validation.md |
/updateprd | Mark complete | validation.md | Updated state.json |
/audit | Local dev tooling | CONVENTIONS.md | Updated tools, report |
/pipeline | CI/CD workflows | CONVENTIONS.md | Workflow files, report |
/infra | Production infra | CONVENTIONS.md | Configs, report |
/documents | Audit/update docs | state.json, PRD, codebase | Updated docs, docs-audit.md |
/carryforward | Capture lessons & security posture | Completed phase artifacts | CONSIDERATIONS.md, SECURITY-COMPLIANCE.md |
/phasebuild | Create new phase | PRD | PRD/phase_NN/ |
Utility Commands
Standalone helpers that operate outside the session workflow. They do not affect session state and can be run at any time. See docs/UTILITIES.md for the full reference and conventions for adding new utility commands.
Additional Resources
Scripts Directory
Utility scripts are available at two locations:
- Plugin:
${CLAUDE_PLUGIN_ROOT}/scripts/(default, always up-to-date) - Local:
.spec_system/scripts/(optional, for per-project customization)
Local scripts take precedence - if .spec_system/scripts/ exists, commands use local scripts instead of plugin scripts.
Available scripts:
analyze-project.sh- Project state analysis (supports--jsonfor structured output)check-prereqs.sh- Environment and tool verification (supports--jsonfor structured output)common.sh- Shared functions
To copy scripts locally during /initspec, choose "copy locally" when prompted. To revert to plugin scripts, delete .spec_system/scripts/.
Hybrid Architecture
Commands use a hybrid approach for reliability:
- Deterministic State (
analyze-project.sh --json): Authoritative state facts - Environment Verification (
check-prereqs.sh --json): Tool and prerequisite validation - Semantic Analysis (Claude): Interprets PRD content, makes recommendations
Why this matters:
- Script output is deterministic - same input always gives same output
- Eliminates risk of Claude misreading state.json
- Tool verification catches missing dependencies BEFORE implementation starts
- Users can run scripts independently to debug
- Claude focuses on what it does best: understanding context and reasoning
analyze-project.sh JSON Output:
{
"project": "project-name",
"monorepo": true,
"packages": [
{"name": "web", "path": "apps/web", "type": "frontend", "stack": "TypeScript + React"}
],
"active_package": {"name": "web", "path": "apps/web"},
"monorepo_detection": null,
"current_phase": 1,
"current_session": "phase01-session02-feature",
"completed_sessions": [
{"id": "phase00-session01-setup", "package": null}
],
"candidate_sessions": [
{"file": "session_01_auth", "completed": false, "package": "apps/web"}
]
}
For single-repo projects: monorepo is null or false, packages is [], active_package is null, and completed_sessions uses the string array format.
check-prereqs.sh JSON Output:
{
"overall": "pass",
"environment": {
"spec_system": {"status": "pass"},
"jq": {"status": "pass", "info": "jq-1.7"}
},
"tools": {
"node": {"status": "pass", "info": "v20.10.0"},
"docker": {"status": "fail", "info": "not installed"}
},
"package": {
"registered": {"status": "pass", "info": "apps/web"},
"directory": {"status": "pass", "info": "apps/web"},
"manifest": {"status": "pass", "info": "package.json"},
"stack": {"status": "pass", "info": "TypeScript"}
},
"workspace": {
"status": {"status": "pass", "info": "monorepo detected"},
"manager": {"status": "pass", "info": "pnpm"},
"runner": {"status": "pass", "info": "turbo"}
},
"database": {
"type": {"status": "pass", "info": "PostgreSQL"},
"migration_tool": {"status": "pass", "info": "prisma"},
"tool_available": {"status": "pass", "info": "npx prisma"},
"seed_script": {"status": "warn", "info": "no seed script found"}
},
"issues": [
{"type": "tool", "name": "docker", "message": "required tool not installed"}
]
}
The package and workspace sections appear only when --package is used or monorepo is detected. The database section only appears when DB signals are detected in the project. For single-repo projects without databases, these sections are empty objects ({}).
Commands and their script usage:
| Command | analyze-project.sh | check-prereqs.sh |
|---|---|---|
/plansession | State + candidates | - |
/implement | Current session | Environment + tools |
/validate | Current session | - |
/documents | State + progress | - |
Monorepo Support
The system supports monorepo projects through a "gentle assumption" -- it auto-detects multi-package structures and offers package-aware workflows, but never requires them. Single-repo projects see zero behavioral change.
How It Works
-
Detection is layered across commands:
/initspec(brownfield): Detects existing workspace configs (pnpm, npm, turbo, nx, cargo, go, lerna)/createprd(greenfield): Parses PRD content for multi-package signals/phasebuild: Checkpoint -- warns if PRD references packages but state lacks them- Later commands: Opportunistic detection if workspace configs appear
-
Package context flows through sessions:
/plansessiondetermines the target package (user input, stub annotation, or prompt)- spec.md headers include
Package:andPackage Stack:fields - Task file paths are scoped to the package directory (e.g.,
apps/web/src/auth.ts) /implementvalidates files stay within declared package scope/updateprdrecords package in completed_sessions
-
Session numbering is global within a phase:
- Sessions interleave across packages (session 02 might be
apps/web, session 03apps/api) - Phase completion requires ALL sessions done, regardless of which packages they target
- One ordered list of sessions per phase -- simple progress tracking
- Sessions interleave across packages (session 02 might be
-
Scripts accept
--packageflag for scoped operations:analyze-project.sh --json --package apps/web-- filters candidates by packagecheck-prereqs.sh --json --package apps/web-- validates package-specific tools
Supported Workspace Managers
| Manager | Detection File |
|---|---|
| pnpm | pnpm-workspace.yaml |
| npm/yarn | package.json with "workspaces" field |
| Turborepo | turbo.json |
| Nx | nx.json |
| Cargo | Cargo.toml with [workspace] |
| Go | go.work |
| Lerna | lerna.json |
Per-Package Stack Detection
Each package gets a stack field based on its manifest files:
| Stack | Indicator |
|---|---|
| TypeScript | tsconfig.json |
| JavaScript | package.json (no tsconfig) |
| Rust | Cargo.toml |
| Go | go.mod |
| Python | pyproject.toml, setup.py, or requirements.txt |
| Ruby | Gemfile |
| Java | pom.xml, build.gradle, or build.gradle.kts |
Best Practices
- Start with /plansession - Always analyze state before choosing work
- One session at a time - Complete before starting next
- MVP first - Defer polish and optimizations
- Validate encoding - Check ASCII before committing
- Update tasks continuously - Mark checkboxes immediately
- Trust the system - Follow workflow, resist scope creep
- Read before implementing - Review spec.md and tasks.md first
- Keep docs current - Run
/documentsafter completing a phase or adding packages
Troubleshooting
| Problem | Solution |
|---|---|
| Scope too large | Split session in PRD before /plansession |
| ASCII validation fails | Run grep -P '[^\x00-\x7F]' to find issues |
| State out of sync | Manually update .spec_system/state.json |
| Commands not found | Verify plugin is enabled |
| Tasks taking too long | Reduce scope, defer non-MVP items |
| Missing tools | Run check-prereqs.sh --tools "tool1,tool2" to verify |
| Environment issues | Run check-prereqs.sh --env to diagnose |
| Stale documentation | Run /documents to audit and update |
| Missing docs | Run /documents to create standard files |
| Lint/format issues | Run /audit to add tooling and auto-fix |
| CI failures | Run /pipeline to add workflows and fix errors |
| Infra not validated | Run /infra to configure health, security, backup, deploy |
| Monorepo not detected | Run /initspec on an existing repo or describe packages in PRD for /createprd |
| Wrong package context | Specify package when running /plansession (e.g., "plan for apps/web") |
| Cross-package session | Set package to null for sessions spanning multiple packages |
| Package tools missing | Run check-prereqs.sh --json --package apps/web to diagnose |
Monorepo Error Recovery
Monorepo not detected by /initspec: Workspace config may have been added after init, or uses a non-standard layout. Manually set "monorepo": true in state.json, add a packages array, add the Workspace Structure table to CONVENTIONS.md, then run /createprd or /plansession to pick up the config.
Session planned for wrong package: If caught before /updateprd, delete the session spec directory and re-run /plansession with the correct package. If caught after /updateprd, manually edit the package field in the relevant completed_sessions entry in state.json.
Package added mid-project: Add the package entry to state.json's packages array, update CONVENTIONS.md's Workspace Structure table, add session stubs to PRD/phase_NN/ with Package: annotations, and increment session_count in the phase's state entry.
Cross-package dependency discovered mid-session: Complete the current session with stubs/interfaces, note the dependency in implementation-notes.md, and plan the shared package session next. Use a cross-cutting session (package: null) when work genuinely spans packages.
Different packages need different tool versions: Use workspace-level version management (.nvmrc per package, engines in package.json). /audit handles shared tools at root with per-package overrides. Validate with check-prereqs.sh --package apps/web.
State disagrees with workspace configs: Run analyze-project.sh --json -- it reports both the state.json monorepo flag and live monorepo_detection. Compare them and update state.json manually if needed.
Decision Quick Reference
| Situation | Action |
|---|---|
| New empty project | /initspec sets monorepo: null, deferred to /createprd |
| Existing monorepo project | /initspec auto-detects, confirms with user |
| PRD describes multiple services | /createprd prompts to confirm monorepo |
| Session targets one package | /plansession scopes spec and tasks to that package |
| Session spans packages | Use cross-cutting session (package: null) |
| Need to add a package mid-project | Manually update state.json and CONVENTIONS.md |
| Phase has mixed-package sessions | Normal -- sessions interleave, phase completes when all done |
| Different stacks per package | /audit and /pipeline handle per-package tool config |
Source
git clone https://github.com/moshehbenavraham/apex-spec-system/blob/main/skills/spec-workflow/SKILL.mdView on GitHub Overview
Apex Spec Workflow provides a specification-driven process that breaks large projects into 2-4 hour sessions with 12-25 tasks each. It organizes work into three stages—Initialization, Sessions, and Phase Transition—anchored in a .spec_system directory to keep specs, PRDs, and phase artifacts in sync. This approach improves focus, traceability, and incremental delivery of AI-assisted development.
How This Skill Works
Technically, the system adheres to a 13-Command Workflow across three stages: Initialization, Sessions, and Phase Transition. You initialize the spec system with /initspec, optionally generate a PRD via /createprd, and create the first phase with /phasebuild. The workflow then loops through /plansession, /implement, /validate, and /updateprd until a phase completes, after which /audit, /pipeline, and /infra support the next phase.
When to Use It
- Setting up and referencing a spec system within a project that uses .spec_system.
- Planning, implementing, validating, and updating sessions (plansession → implement → validate → updateprd).
- Generating or manually populating PRDs from requirements or design docs (PRD/UX PRD).
- Creating the first phase structure with /phasebuild and looping through sessions until the phase is complete.
- Executing phase transitions and infrastructure steps with /audit, /pipeline, and /infra to prepare for the next phase.
Quick Start
- Step 1: /initspec to set up the spec system in your project.
- Step 2: /createprd or manually populate PRD with requirements; optionally generate UX PRD.
- Step 3: /phasebuild to lay out the first phase, then /plansession to begin the first session.
Best Practices
- Always run /initspec first to establish the spec system in the project.
- Keep sessions short and scoped: 2-4 hours with 12-25 tasks per the Core Philosophy.
- Maintain a clear mapping between specs in .spec_system and the PRD/UX PRD artifacts.
- Use /validate after each session to confirm completeness before /updateprd.
- Document lessons learned with /carryforward and keep the directory tidy (state.json, specs, PRD).
Example Use Cases
- New feature in a monorepo: initialize with /initspec, generate PRD via /createprd, and create phase structure with /phasebuild.
- Plan a session with /plansession, have AI drive /implement task-by-task, then verify with /validate.
- Update PRD after a session with /updateprd to reflect completed tasks.
- Run a full Phase Transition: /audit → /pipeline → /infra → /carryforward.
- Work within a project using a single .spec_system at repo root for monorepos.