Skip to content
/

spec_driven_develop

zhu1090093659-sdd · zhu1090093659/spec_driven_develop · ★ 866 · last commit 2026-05-25

6-phase cybernetics-inspired pipeline for large-scale AI transformations with S.U.P.E.R architectural health framework, adaptive drift control, and GitHub-native task tracking.

Best whenAI development plans should self-correct via closed-loop feedback control (Qian Xuesen's engineering cybernetics), not blindly execute when reality diverges …
Skip ifProceeding when information is insufficient (Deep Discuss Phase 2 gate), Marking tasks complete without passing S.U.P.E.R Quick Check
vs seeds
superpowers(Archetype 1 — skills-only with multi-agent dispatch) in orchestration model, but significantly more sophisticated: name…
Primitive shape 5 total
Skills 2 Subagents 3
00

Summary

zhu1090093659/spec_driven_develop — Summary

A platform-agnostic Claude plugin shipping two complementary skills: Spec-Driven Develop (6-phase pipeline for large-scale transformations: analysis → intent refinement → task decomposition → progress tracking → execution → archive) and Deep Discuss (7-phase structured discussion for problem analysis and solution design). Both are pure Markdown — no SDK, no runtime, no external dependencies.

The framework's distinguishing feature is the S.U.P.E.R architectural framework (Single Purpose, Unidirectional Flow, Ports over Implementation, Environment-Agnostic, Replaceable Parts) woven into every phase: analysis rates each module's S.U.P.E.R compliance, planning prioritizes fixing violations, and execution includes a S.U.P.E.R Quick Check before marking any task complete.

Version 1.11.0 adds GitHub-native task tracking (Issues, Milestones, Projects) with three auto-detected modes (GITHUB_FULL / GITHUB_STANDARD / LOCAL_ONLY) and an adaptive control loop inspired by engineering cybernetics: drift thresholds at 20%/40%/60% of tasks trigger automatic re-planning or scope re-evaluation.

Three named subagents (project-analyzer, task-architect, task-executor) handle parallel analysis and decomposition. 866 stars — by far the highest-starred framework in this batch.

01

Overview

zhu1090093659/spec_driven_develop — Overview

Origin

Created by zhu1090093659. Repository description: "A structured development methodology for AI coding agents, inspired by Qian Xuesen's engineering cybernetics. Closed-loop feedback control keeps plans aligned with reality. Pure Markdown, any platform." 866 stars, 92 forks. Active — version 1.11.0, last commit 2026-05-25.

Philosophy

The framework is explicitly inspired by Qian Xuesen's engineering cybernetics (工程控制论) — the idea that a system should observe its actual state, compare it to the desired state, and automatically correct deviations. Applied to AI-assisted development: instead of blindly executing a plan, the workflow collects execution telemetry after every task and corrects course when drift exceeds thresholds.

From the README:

"Spec-Driven Develop is a platform-agnostic AI agent plugin that ships two complementary skills: Spec-Driven Develop — automates the full development pipeline for large-scale complex tasks: deep project analysis, phased task decomposition, document-driven progress tracking, and execution — all within a single session."

On the S.U.P.E.R framework:

"Write code like building with LEGO — each brick has a single job, a standard interface, a clear direction, runs anywhere, and can be swapped at will."

On Deep Discuss:

"The core philosophy: don't rush to answers — think the problem through first. Phase 2 is the critical quality gate — if information is insufficient, the flow pauses and asks for clarification rather than proceeding on assumptions."

The adaptive control loop (v1.10 addition)

Inspired by cybernetics, the controller observes execution reality and corrects:

  • Mild drift (≥ 20% of phase tasks deviate): annotate next task with warning
  • Significant drift (≥ 40%): halt and re-decompose remaining tasks
  • Severe drift (≥ 60%): return to Phase 2 for scope re-evaluation

This ensures the workflow self-corrects instead of blindly executing an invalidated plan.

02

Architecture

zhu1090093659/spec_driven_develop — Architecture

Distribution

  • Type: claude-plugin + standalone-repo (platform-agnostic)
  • Install: Claude Code plugin marketplace, Codex plugin, or manual copy of SKILL.md
  • Install complexity: multi-step

Install commands

# Claude Code
/plugin marketplace add zhu1090093659/spec_driven_develop
/plugin install spec-driven-develop@spec-driven-develop
/reload-plugins

# Codex
codex plugin marketplace add zhu1090093659/spec_driven_develop --ref main

# Cursor / any other agent
# Copy plugins/spec-driven-develop/skills/spec-driven-develop/SKILL.md
# to wherever your agent reads instructions

Directory tree

spec_driven_develop/
├── .agents/
│   └── plugins/
│       └── marketplace.json      ← Agent marketplace metadata
├── .claude-plugin/
│   └── marketplace.json          ← Claude Code marketplace metadata
├── plugins/
│   └── spec-driven-develop/
│       ├── .claude-plugin/
│       │   └── plugin.json       ← Plugin manifest
│       ├── .codex-plugin/
│       │   └── plugin.json       ← Codex plugin manifest
│       ├── skills/
│       │   ├── spec-driven-develop/
│       │   │   └── SKILL.md      ← Main 6-phase pipeline skill
│       │   └── deep-discuss/
│       │       └── SKILL.md      ← 7-phase discussion skill
│       ├── agents/
│       │   ├── project-analyzer.md   ← Analysis subagent
│       │   ├── task-architect.md     ← Decomposition subagent
│       │   └── task-executor.md      ← Execution subagent
│       └── commands/
├── docs/
│   └── archives/
├── README.md
└── README.zh-CN.md

Target AI tools

Platform-agnostic. Explicitly tested on:

  • Claude Code (first-class with plugin and enhanced agent support)
  • Codex (CLI plugin)
  • Cursor (global/project-level skill)

Works with any agent that reads Markdown: Windsurf, Cline, Aider, Continue, Roo Code, Augment.

Required runtime

None. Pure Markdown.

03

Components

zhu1090093659/spec_driven_develop — Components

Skills (2)

Skill File Purpose
spec-driven-develop plugins/spec-driven-develop/skills/spec-driven-develop/SKILL.md 6-phase pipeline for large-scale transformations
deep-discuss plugins/spec-driven-develop/skills/deep-discuss/SKILL.md 7-phase structured discussion for problem analysis

Subagents (3, persona-md format)

Agent File Purpose
project-analyzer agents/project-analyzer.md Phase 1 parallel analysis (architecture, module inventory, risk/SUPER health)
task-architect agents/task-architect.md Phase 3 task decomposition (2 parallel strategies if large enough)
task-executor agents/task-executor.md Phase 5 task execution (isolated worktree + PR per task)

Commands

None.

Hooks

None.

MCP Servers

None bundled.

Scripts

None.

Templates

Referenced in references/templates/ within the skill:

  • references/templates/analysis.md — Phase 1 analysis doc templates
  • references/templates/plan.md — Phase 3 planning doc templates
  • references/templates/progress.md — Phase 4 progress/MASTER.md templates

Reference files (loaded by skill as needed)

File Content
references/super-philosophy.md S.U.P.E.R principles with examples
references/behavioral-rules.md Agent behavioral constraints
references/parallel-protocol.md Parallel execution protocol
references/github-integration.md GitHub Issues/Milestones/Projects protocol
references/adaptive-control.md Drift threshold and control loop protocol

S.U.P.E.R framework (woven into every phase)

Principle Acronym How Enforced
Single Purpose S Module inventory rates SP compliance per module
Unidirectional Flow U Architecture health check flags circular dependencies
Ports over Implementation P Module inventory evaluates schema-defined I/O
Environment-Agnostic E Risk assessment catches hardcoded config
Replaceable Parts R Replacement cost rated per module
05

Prompts

zhu1090093659/spec_driven_develop — Prompts

Excerpt 1 — spec-driven-develop SKILL.md: Phase 1 parallel dispatch (verbatim)

1. Launch `project-analyzer` sub-agents **in parallel** to analyze the codebase concurrently. Split the work by focus area:
   - **Agent 1 — Architecture & Stack**: Project structure, directory layout, technology stack, entry points, build/run commands
   - **Agent 2 — Module Inventory**: Each module's responsibility, public API surface, size, internal/external dependencies. **Must evaluate each module against all five S.U.P.E.R principles** (Single Purpose, Unidirectional Flow, Ports over Implementation, Environment-Agnostic, Replaceable Parts) and assign a per-principle compliance rating.
   - **Agent 3 — Risks & S.U.P.E.R Health**: Transformation risks, complexity hotspots, platform-specific code, coding conventions. **Must produce a S.U.P.E.R Architecture Health Summary** evaluating the overall codebase against each principle, identifying violation hotspots that become priority targets in the transformation plan.

   Provide each agent with the preliminary direction from Phase 0 AND `references/super-philosophy.md` so they can assess findings against S.U.P.E.R principles in context of the intended transformation.

   If sub-agents are not available on the current platform, perform the analysis sequentially yourself — the scope is the same either way.

Prompting technique: Graceful degradation instruction. The skill is designed for multi-agent platforms but explicitly handles the single-agent case ("perform the analysis sequentially yourself"). This makes the skill portable without breaking on platforms without subagent support.

Excerpt 2 — S.U.P.E.R Quick Check (verbatim from SKILL.md)

S.U.P.E.R Code Review Checklist (run after every task before marking complete):

| Check | Principle |
|:------|:----------|
| Every new module/file has exactly one responsibility | S |
| No function does more than one conceptual thing | S |
| Data flows input → processing → output, no reverse deps | U |
| No circular imports introduced | U |
| Cross-module interfaces are schema-defined | P |
| Module I/O is serializable | P |
| No hardcoded paths, URLs, keys, or config values | E |
| All new dependencies explicitly declared | E |
| New modules can be replaced without changes to others | R |
| All tests pass after the change | — |

**All pass = proceed. 1-2 fail = fix before marking complete. 3+ fail = stop and refactor.**

Prompting technique: Gated checklist with explicit numeric thresholds. The 1-2 fail vs. 3+ fail distinction creates a decision boundary rather than leaving it to agent discretion, reducing inconsistent enforcement.

Excerpt 3 — Adaptive control initialization (verbatim from SKILL.md)

For each Milestone created, compute the percentage-based drift thresholds from the task count in that phase and append the adaptive control YAML block to the Milestone description:
```yaml
---
adaptive:
  drift_score: 0
  strategy: "<decomposition-strategy>"
  thresholds:
    annotate: <ceil(total_tasks * 0.20)>
    replan: <ceil(total_tasks * 0.40)>
    rescope: <ceil(total_tasks * 0.60)>
  total_tasks: <count>
  completed_tasks: 0
  last_updated: "<ISO-8601>"

**Prompting technique**: Computed thresholds with ceiling function. Rather than fixed numbers, thresholds are calculated from the actual task count — making the adaptive control system scale correctly regardless of plan size.
09

Uniqueness

zhu1090093659/spec_driven_develop — Uniqueness

differs_from_seeds

This framework does not fit cleanly into any single seed archetype. It is closest to superpowers (Archetype 1 — skills-only behavioral framework with multi-agent dispatch) in its orchestration model — the SKILL.md dispatches named subagents (project-analyzer, task-architect, task-executor) in parallel phases, and tasks execute in isolated worktrees. However, it exceeds superpowers' multi-agent sophistication: superpowers uses the general-purpose Task tool to spawn skill chains, while spec_driven_develop dispatches named persona-md agents with specific role assignments. The adaptive control loop (drift thresholds triggering re-planning or rescoping, inspired by Qian Xuesen's cybernetics) is entirely absent from all 11 seeds. The S.U.P.E.R architectural health framework woven into every phase (per-module compliance ratings, S.U.P.E.R Quick Check as a task gate) is also unique — no seed enforces an architectural philosophy at this granularity. The GitHub-native task tracking (Issues + Milestones + Projects, auto-detected from gh CLI scope) resembles taskmaster-ai's task management approach (Archetype 3 — MCP-anchored) but without a bundled MCP server.

Positioning

The highest-starred single-author SDD framework in this batch (866 stars, 92 forks). Targets large-scale transformation projects (rewrites, migrations, architecture overhauls) rather than feature-by-feature development. The "Spec-Driven" name is somewhat misleading — this is more of a project management and execution framework that happens to start with analysis and planning rather than a spec-writing tool.

Observable failure modes

  1. Scoped for large projects: The 6-phase pipeline has significant overhead for small features. Triggering on "rewrite" or "migrate" is intentional, but users expecting a lightweight workflow will be disappointed.
  2. GitHub dependency for best experience: LOCAL_ONLY mode lacks the durability and collaboration features of GitHub integration. Teams without gh CLI access or GitHub repos lose significant value.
  3. S.U.P.E.R checker is agent self-assessment: The S.U.P.E.R Quick Check is a prompt checklist run by the agent, not a programmatic linter. Enforcement quality depends on the agent following it faithfully.
  4. Long single-session design: The framework is designed to complete in one long session. Very large projects may hit context limits before Phase 6, despite MASTER.md resumption support.
  5. Subagent platform dependency: Phase 1 parallel analysis assumes three simultaneous subagents. On platforms without subagent support, sequential fallback adds significant time.
04

Workflow

zhu1090093659/spec_driven_develop — Workflow

Spec-Driven Develop: 6-phase pipeline

Phase Name Actions Artifact
0 Quick Intent Capture Extract direction (1-2 sentences), confirm with user Direction statement
1 Deep Project Analysis Launch 3 parallel project-analyzer agents (architecture, module inventory, risk+SUPER health); GitHub pre-flight docs/analysis/project-overview.md, module-inventory.md, risk-assessment.md
2 Intent Refinement Present Phase 1 findings, ask targeted questions grounded in analysis, confirm scope, S.U.P.E.R priorities Confirmed task definition
3 Task Decomposition Launch task-architect agents (parallel strategies); create GitHub Issues/Milestones/Labels; initialize adaptive control state docs/plan/task-breakdown.md, dependency-graph.md, milestones.md
4 Progress Tracking Generate MASTER.md (GitHub index or local tracker) docs/progress/MASTER.md
5 Confirm & Execute Present summary, get confirmation, execute via task-executor agents (parallel lanes, worktrees, PRs); adaptive control loop code + PR per task
6 Archive Preserve all artifacts docs/archives/<project>/

Cross-conversation continuity

Critical: Check for docs/progress/MASTER.md before starting.

  • If exists: resume from last checkpoint (query GitHub for merged Issues if GitHub mode)
  • If not: fresh start from Phase 0

GitHub integration modes (auto-detected)

Mode Requirements What's created
GITHUB_FULL gh CLI + auth + project scope Issues + Milestones + Labels + Project board + worktree + PR
GITHUB_STANDARD gh CLI + auth + repo scope Issues + Milestones + Labels + worktree + PR (no board)
LOCAL_ONLY None Local Markdown workflow

Adaptive control loop

After every task, agent collects telemetry and updates drift_score:

  • Mild (≥ 20%): annotate next task
  • Significant (≥ 40%): halt + re-decompose remaining
  • Severe (≥ 60%): return to Phase 2 for scope re-evaluation

Deep Discuss: 7-phase flow

Phase Purpose
1. Receive Information Listen, restate, confirm understanding
2. Problem Audit Validate problem, check info sufficiency, surface hidden issues (critical quality gate)
3. Deep Analysis Multi-angle root cause analysis with confidence levels
4. Solution Design 2-3 options with trade-offs and recommendations
5. Self-Review First review of proposed solutions
6. Final Review Completeness check, risk mitigation, verification plan
7. Execution Only when user explicitly says "go"

Approval gates

Gate Type
Phase 2: Explicit user confirmation before Phase 3 yes-no
Phase 5: Summary confirmation before execution yes-no
Deep Discuss Phase 7: "go" keyword required typed-confirm
Deep Discuss Phase 2: Information sufficiency check (pause and ask if insufficient) freetext-clarify
06

Memory Context

zhu1090093659/spec_driven_develop — Memory & Context

State storage

File-based with optional GitHub integration:

Local files:

docs/
  analysis/
    project-overview.md
    module-inventory.md
    risk-assessment.md
  plan/
    task-breakdown.md
    dependency-graph.md
    milestones.md
  progress/
    MASTER.md              ← Cross-session continuity anchor
  archives/<project>/      ← Phase 6 archive

GitHub state (GITHUB_FULL / GITHUB_STANDARD modes):

  • Issues (one per task) — auto-closed on PR merge
  • Milestones (one per phase) — with adaptive control YAML in description
  • Labels (priority, size, phase, lane, spec-driven)
  • Project board (GITHUB_FULL only)

Cross-session continuity

MASTER.md is the anchor:

  • Skill checks for MASTER.md before any action
  • If exists → resume from last checkpoint (read tracking mode, query GitHub for merged Issues if GitHub mode)
  • If not exists → fresh start from Phase 0

From the SKILL.md:

"CRITICAL: Before starting any phase, check if docs/progress/MASTER.md already exists in the project."

Adaptive control state (per Milestone)

YAML appended to each GitHub Milestone description:

adaptive:
  drift_score: 0
  thresholds: {annotate: N, replan: N, rescope: N}
  total_tasks: N
  completed_tasks: 0

In LOCAL_ONLY mode, this state lives in MASTER.md.

Compaction handling

Not explicitly addressed by hooks. The MASTER.md continuity check handles cross-session recovery naturally.

Memory type

Hybrid: file-based (primary) + GitHub (optional tracking layer)

Persistence

Project-level (docs/) + global-accessible (GitHub Issues/Milestones)

07

Orchestration

zhu1090093659/spec_driven_develop — Orchestration

Multi-agent support

Yes. Three named subagents with parallel dispatch:

  • Phase 1: 3 project-analyzer agents dispatched in parallel (architecture, module inventory, risks)
  • Phase 3: 1-2 task-architect agents (parallel strategies for large projects)
  • Phase 5: Multiple task-executor agents in parallel lanes (tasks with no mutual dependencies)

Orchestration pattern

Hierarchical (orchestrator skill dispatches specialist subagents). The spec-driven-develop SKILL.md is the conductor; it dispatches project-analyzer, task-architect, and task-executor as workers.

Isolation mechanism

Git worktree per task (in Phase 5 execution). Each task-executor works in an isolated worktree, creates a PR linked to its GitHub Issue (closes #N), and the Issue auto-closes on merge.

Execution mode

One-shot per session (but with cross-session continuity via MASTER.md). The skill is designed to complete the entire pipeline within a single long session, with MASTER.md enabling resumption.

Multi-model routing

None. Model-agnostic.

Prompt chaining

Yes, explicit. Phase outputs feed subsequent phases:

  • Phase 0 direction → Phase 1 analysis focus
  • Phase 1 S.U.P.E.R health → Phase 2 targeted questions
  • Phase 2 confirmed definition → Phase 3 task decomposition
  • Phase 3 task breakdown → Phase 4 MASTER.md
  • Phase 4 MASTER.md → Phase 5 execution

Consensus mechanism

None. The orchestrator skill makes all decomposition decisions; subagents execute within their assigned scope.

Parallel execution

Yes, in two places:

  1. Phase 1: 3 analysis agents in parallel
  2. Phase 5: task execution in parallel lanes (tasks with no mutual dependencies)

Adaptive control loop

After every task execution, drift_score is updated:

  • Actual effort vs. estimated
  • S.U.P.E.R compliance delta
  • Unplanned dependencies

Thresholds: 20% = annotate, 40% = replan, 60% = rescope to Phase 2.

Auto-validators

  • S.U.P.E.R Quick Check after every task (before marking complete)
  • Drift score telemetry after every task
  • GitHub pre-flight check (Phase 1) to determine tracking mode
08

Ui Cli Surface

zhu1090093659/spec_driven_develop — UI / CLI Surface

Dedicated CLI binary

None. The framework is pure Markdown skills.

Local web dashboard

None.

IDE integration

Claude Code plugin (primary). After installation:

  • spec-driven-develop skill is available for large-scale transformation requests
  • deep-discuss skill triggers on problem analysis requests

Codex plugin (secondary). Cursor: manual SKILL.md copy.

Platform-specific trigger keywords

From SKILL.md trigger descriptions:

spec-driven-develop triggers on:

  • "rewrite", "migrate", "overhaul", "refactor entire project", "transform", "rebuild in [language]", "spec-driven"
  • Chinese keywords: "改造", "重写", "迁移", "重构", "大规模", "规范驱动"

deep-discuss triggers on:

  • "let's discuss", "help me analyze", "I'm stuck on a decision", or describing a complex problem

GitHub integration surface

When gh CLI is available with appropriate scope:

  • Creates GitHub Issues (one per task)
  • Creates Milestones (one per phase)
  • Creates Labels (priority, size, phase, lane, spec-driven)
  • Creates Project board (GITHUB_FULL mode)
  • Auto-closes Issues via PR merge

Observability

  • MASTER.md — durable progress index (works as audit log)
  • GitHub Issues/Milestones — distributed tracking
  • Adaptive control YAML in Milestone descriptions — drift telemetry
  • Phase archives in docs/archives/ — post-project record

Graceful degradation

The skill explicitly handles platform capability variations:

  • No subagents → sequential analysis
  • No gh CLI → LOCAL_ONLY mode
  • No Project board scope → GITHUB_STANDARD mode

Related frameworks

same archetype · same primary tool · same memory type

claude-mem (thedotmack) ★ 78k

Background worker service captures every tool call as an observation, AI-compresses sessions, and auto-injects relevant past…

pi (badlogic/earendil) ★ 55k

A minimal, hackable, multi-provider terminal coding agent that adapts to your workflows via npm-installable TypeScript Extensions…

Agent Skills (Addy Osmani) ★ 46k

Encodes senior-engineer software development lifecycle as 23 auto-routed skills and 7 slash commands for any AI coding agent.

wshobson/agents Plugin Marketplace ★ 36k

Single Markdown source for 83 domain-specialized plugins that auto-generates idiomatic artifacts for five AI coding harnesses.

TabbyML/Tabby ★ 34k

Self-hosted AI coding assistant server (alternative to GitHub Copilot) with admin dashboard, RAG-based completions, and multi-IDE…

Compound Engineering ★ 17k

Make each unit of engineering work compound into easier future work via brainstorm→plan→execute→review→learn cycles.