Skip to content

rccypher/agentic-dev-framework

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Agentic Development Framework

A 16-step structured development process for AI coding agents. Designed to reduce rework, prevent repeated mistakes, and ensure what gets built matches what was asked for.

The Problem

AI coding agents are powerful but fail in predictable ways:

  • They lose context across sessions and repeat the same mistakes
  • They start coding before confirming what the user actually wants
  • They verify code works but not that it meets requirements
  • They don't learn from debugging dead ends
  • They introduce inconsistent patterns because they don't check existing code
  • They deploy stale builds and debug phantom bugs
  • They expand scope to fill available compute
  • They don't think about what could go wrong before building
  • They treat security as an afterthought

Every step in this framework exists because its absence caused a specific, documented failure.

Built From

Validated against 22+ engineering frameworks:

Category Frameworks
AI Agent Workflows BMAD, GitHub Spec Kit, Amazon Kiro, OpenSpec, RIPER-5, Memory Bank, spec-then-code, Aider
Professional Engineering Google Engineering Practices, Microsoft SDL, Amazon ORR/COE, Meta DRS, DORA
Risk Management Risk Up Front (RUF), Risk Storming, ATAM, Risk-First, Pre-Mortem Analysis
Development Methods XP (TDD), Shape Up, Trunk-Based Development, Toyota/Lean
Reliability Google SRE, Netflix Chaos Engineering, Blameless Postmortems
Documentation ADRs, RFC Culture (Uber, HashiCorp, Google Design Docs)

Quick Start

1. Copy into your project

git clone https://github.com/rccypher/agentic-dev-framework.git
cp agentic-dev-framework/FRAMEWORK.md your-project/Docs/DEVELOPMENT_PROCESS.md
cp agentic-dev-framework/DISCOVERY.md your-project/Docs/DISCOVERY.md
cp -r agentic-dev-framework/templates your-project/Docs/templates
mkdir -p your-project/Docs/{archive,design-discussions,specs}

2. Add to your CLAUDE.md

See install/CLAUDE_SNIPPET.md for the snippet.

3. Use it

The framework routes tasks through 3 complexity tiers:

Tier When Steps
Quick Fix Single file, obvious fix 5 steps
Standard 2-3 files, clear scope 11 steps
Complex 3+ modules, architectural All 16 steps

The 16 Steps

 1. COMPLEXITY ASSESSMENT      Route + time budget + task type
 2. DESIGN DISCUSSION LOG      Discovery process + structured elicitation
 3. REQUIREMENTS CONFIRMATION  Restate, acceptance criteria, user confirms
 4. SPEC ARTIFACT              Formal spec with non-goals + pre-mortem + security
 5. EXPLORE                    Read code, identify modules, note patterns
 6. READ ARCHIVE               Invariants, dead ends, constraints, postmortems
 7. IMPACT ANALYSIS            Blast radius + risk storming + riskiest assumption
 8. PLAN                       Approach + alternatives + RAT + pre-mortem + tests + rollback
 9. WRITE TESTS                Failing tests before implementation (red phase)
10. IMPLEMENT                  Pattern conformance + scope completeness + Andon stops
11. VERIFY                     Build, deploy, tests green, ORR, self-review checklist
12. DEBUG                      Root cause + five whys + record dead ends in real time
13. REQUIREMENTS VERIFICATION  Walk spec with evidence, not assertions
14. FRESH-AGENT REVIEW         New session reviews spec + diff (zero conversation history)
15. ARCHIVE CAPTURE            Decisions, dead ends, prompt rationale, process reflection
16. DONE                       Checklist verification

Key Files

File Purpose
FRAMEWORK.md Full 16-step process with all details
DISCOVERY.md Non-technical user elicitation guide with metaphors
templates/spec.md Spec template with non-goals, pre-mortem, STRIDE
templates/design-discussion.md Design discussion with discovery and DKDK
templates/archive-entry.md Archive entry format and guidelines
templates/coe-postmortem.md Incident analysis with Five Whys and feedback loop
install/CLAUDE_SNIPPET.md CLAUDE.md integration snippet

What Makes This Different

Unique to this framework (not found in any surveyed methodology):

  • Dead-end recording during active debugging
  • Prompt/LLM rationale capture in the archive
  • Scope boundary confirmation as an explicit gate
  • Pre-plan archive review for known failures
  • Blast radius + caller/callee tracing before code changes
  • Rollback path required in every plan
  • Non-technical discovery process with practical metaphors
  • Andon cord (hard stop conditions) during implementation
  • COE feedback loop from incidents back into the framework itself

Adopted and adapted from professional engineering:

  • Pre-mortem analysis (Gary Klein) → in spec and plan
  • STRIDE threat modeling (Microsoft SDL) → in spec
  • Operational Readiness Review (Amazon ORR) → in verify
  • Five Whys root cause analysis (Amazon COE) → in debug and postmortems
  • Self-review checklist (Google Engineering) → in verify
  • Circuit breaker and time budgets (Shape Up) → in complexity assessment
  • Test-first gates (XP TDD) → write tests step
  • Fresh-agent review (OpenSpec) → catch implementation rationalizations
  • Error budget tracking (Google SRE) → ongoing quality metric
  • DKDK meetings for unknown unknowns (Risk Up Front) → in design discussion
  • Riskiest Assumption Test (Lean) → in impact analysis and plan
  • Risk Storming (Simon Brown) → in impact analysis
  • Feature flags (Trunk-Based Development) → in implement

v3.0 additions — Agent Team Structure & Prompting (empirical):

  • Single-agent default over multi-agent (Cursor, Windsurf, Copilot convergence)
  • Artifact-driven communication, not conversational (MetaGPT 3.9 vs ChatDev 2.1)
  • Model tiering: reasoning models for design/debug, standard for implementation
  • Fresh context per implementation unit (BMAD, validated in production)
  • No persona framing in prompts — domain checklists + methodology instead (arXiv:2603.18507: expert personas degrade coding accuracy by -0.65 on MT-Bench)
  • Adversarial review framing with mandatory findings (BMAD TEA agent pattern)
  • Tool permissions enforce role boundaries, not prompt instructions

Task-Type Variants

Type Key Differences
Feature Full pipeline (default)
Bug Fix Current vs expected vs unchanged behavior
Refactoring Tests pass before AND after; no new features
Security Fix STRIDE mandatory; fresh-agent review mandatory; ship immediately
Prompt Engineering A/B comparison on known inputs; regression testing
Infrastructure ORR mandatory; rollback verification; canary deployment

Contributing

This framework evolves through use. Process reflections (step 15) identify friction. Open an issue or PR with improvements.

License

MIT

About

A 16-step development process framework for AI coding agents — structured workflow with complexity routing, spec artifacts, test-first gates, fresh-agent review, and decision archives

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors