Skip to content

Instantly share code, notes, and snippets.

@AlexMikhalev
Created January 20, 2026 23:23
Show Gist options
  • Select an option

  • Save AlexMikhalev/b77cb9e954ff8559981b1852d7694f9c to your computer and use it in GitHub Desktop.

Select an option

Save AlexMikhalev/b77cb9e954ff8559981b1852d7694f9c to your computer and use it in GitHub Desktop.
Terraphim Skills: A Disciplined Development Framework for AI-Assisted Engineering

Terraphim Skills: A Disciplined Development Framework for AI-Assisted Engineering

TL;DR: Terraphim Skills implements a V-model development process as Claude Code skills, combining Greg McKeown's Essentialism philosophy with systematic quality gates. The workflow flows from session start through left-side planning, execution orchestration, right-side verification, and concludes with structured handover.


The Problem with AI-Assisted Development

AI coding assistants are powerful, but power without discipline leads to chaos. Common issues include:

  • Scope creep: AI eagerly adds features nobody asked for
  • Missing requirements: Jumping to code before understanding the problem
  • No traceability: Can't trace code back to requirements
  • Heroic effort: Pushing through complexity instead of simplifying
  • Knowledge loss: Session ends, context disappears

Terraphim Skills addresses these with a structured framework based on two complementary philosophies from Greg McKeown:

  1. Essentialism: The disciplined pursuit of less - focus on the vital few
  2. Effortless: Making essential work easy - find the simpler path

The V-Model: A Visual Overview

graph TB
    subgraph LEFT["LEFT SIDE (Planning)"]
        P1[Phase 1: Research<br/>Problem understanding<br/>Constraints & risks<br/>Success criteria]
        P2[Phase 2: Design<br/>File changes<br/>API signatures<br/>Test strategy]
        P25[Phase 2.5: Specification<br/>Deep interview<br/>Edge cases<br/>Tradeoffs]
    end

    subgraph MIDDLE["EXECUTION"]
        P3[Phase 3: Implementation<br/>Test first<br/>Small commits<br/>Quality checks]
    end

    subgraph RIGHT["RIGHT SIDE (Testing)"]
        P4[Phase 4: Verification<br/>Unit testing<br/>Integration testing<br/>Traceability matrix]
        P5[Phase 5: Validation<br/>System testing<br/>UAT with stakeholders<br/>Formal sign-off]
    end

    P1 --> P2
    P2 --> P25
    P25 --> P3
    P3 --> P4
    P4 --> P5

    P1 -.->|validates| P5
    P2 -.->|verifies| P4

    style LEFT fill:#e1f5fe
    style MIDDLE fill:#fff3e0
    style RIGHT fill:#e8f5e9
Loading

The dotted arrows show traceability: Phase 4 verifies Phase 2 design was built correctly; Phase 5 validates Phase 1 requirements were actually solved.


The Complete Workflow

1. Session Start

Every disciplined session begins with explicit context establishment:

/session-start

This skill:

  • Creates a tracked development session
  • Establishes the working context (branch, recent commits)
  • Links to relevant GitHub issues
  • Sets up auto-documentation for session progress

Why it matters: Without explicit session boundaries, work bleeds across contexts. Session start creates a clean slate with clear scope.


2. Left Side of V: Planning Phases

The left-side-of-v agent orchestrates research and design before any code is written.

flowchart TD
    START([Start Planning]) --> R1

    subgraph RESEARCH["Phase 1: Research (EXPLORE)"]
        R1[Execute disciplined-research skill]
        R2[Create Research Document]
        R3[Apply Essentialism gates]
        R4[Request human approval]
        R1 --> R2 --> R3 --> R4
    end

    R4 --> QE1

    subgraph EVAL1["Phase 1.5: Quality Evaluation"]
        QE1[Score with KLS 6-dimension framework]
        QE2{Min 3.5 avg?}
        QE1 --> QE2
    end

    QE2 -->|NO| R1
    QE2 -->|YES| D1

    subgraph DESIGN["Phase 2: Design (ELIMINATE)"]
        D1[Execute disciplined-design skill]
        D2[Create Implementation Plan]
        D3[Apply 5/25 Rule]
        D4[Document eliminations]
        D5[Request human approval]
        D1 --> D2 --> D3 --> D4 --> D5
    end

    D5 --> S1

    subgraph SPEC["Phase 2.5: Specification"]
        S1[Deep interview]
        S2[Edge cases & tradeoffs]
        S1 --> S2
    end

    S2 --> QE3

    subgraph EVAL2["Phase 2.5: Quality Evaluation"]
        QE3[Score with KLS framework]
        QE4{Min 3.5 avg?}
        QE3 --> QE4
    end

    QE4 -->|NO| D1
    QE4 -->|YES| READY([Ready for Implementation])

    style RESEARCH fill:#e3f2fd
    style DESIGN fill:#e3f2fd
    style SPEC fill:#e3f2fd
    style EVAL1 fill:#fff8e1
    style EVAL2 fill:#fff8e1
Loading

Phase 1: Research (EXPLORE)

Uses the disciplined-research skill to produce a Research Document answering:

  • What problem are we solving?
  • What's the current state?
  • What constraints exist (technical, business, integration)?
  • What are the risks?
  • What questions remain open?

Essentialism Integration:

  • Essential Questions Check: Does this energize? Leverage strengths? Meet a real need?
  • Vital Few: Identify maximum 3 essential constraints
  • 5/25 Rule: Document what's explicitly NOT being investigated

Phase 1.5: Quality Evaluation

Before proceeding to design, the disciplined-quality-evaluation skill scores the research using the KLS 6-dimension framework:

Dimension Question
Syntactic Is it well-formed and complete?
Semantic Does it accurately represent the domain?
Pragmatic Is it useful for the intended purpose?
Social Will stakeholders understand and accept it?
Organizational Does it fit the organizational context?
Technical Is it technically feasible and sound?

Gate: Minimum 3.0 per dimension, 3.5 average required to proceed.

Phase 2: Design (ELIMINATE)

Uses the disciplined-design skill to produce an Implementation Plan:

  • Exact files to change
  • Function signatures and types
  • Step-by-step implementation sequence
  • Test strategy for each step
  • Eliminated options (what we decided NOT to do)

Essentialism Integration:

  • 5/25 Rule: Circle top 5 features, remaining 20 become "Avoid At All Cost"
  • Simplicity Check: "What if this could be easy?"
  • Elimination Mandate: Document what we're removing, not just adding

Phase 2.5: Specification Interview

Optional but powerful: the disciplined-specification skill conducts a structured interview to probe:

  • Edge cases not covered in design
  • Tradeoffs between approaches
  • Hidden requirements
  • Error handling scenarios
  • Performance expectations

Output: Specification Interview Findings appended to the design document.


3. Execution Orchestration: Implementation Phase

The execution-orchestrator agent manages Phase 3 with integrated quality assurance.

flowchart TD
    START([Approved Plan]) --> I1

    subgraph IMPL["Phase 3: Implementation (EXECUTE)"]
        I1[Follow plan step by step]
        I2[Write tests at each step]
        I3[Log friction points]
        I4[Commit at each step]
        I1 --> I2 --> I3 --> I4
    end

    I4 --> Q1

    subgraph QEVAL["Phase 3.5: Quality Evaluation"]
        Q1[Review implementation docs]
        Q2[Analyze Effortless Execution Log]
        Q3[Check for plan deviations]
        Q1 --> Q2 --> Q3
    end

    Q3 --> G1

    subgraph GATE["Quality Gate"]
        G1[Code review]
        G2[UBS Scanner - 1000+ patterns]
        G3[Security audit - if applicable]
        G4[Performance check - if applicable]
        G5{GO / NO-GO}
        G1 --> G2 --> G3 --> G4 --> G5
    end

    G5 -->|NO-GO| I1
    G5 -->|GO| READY([Ready for Verification])

    style IMPL fill:#fff3e0
    style QEVAL fill:#fff8e1
    style GATE fill:#ffebee
Loading

Phase 3: Implementation (EXECUTE)

Uses the disciplined-implementation skill to:

  • Follow the approved plan exactly (no scope creep)
  • Write tests at each step (test-first)
  • Make one commit per step (reviewable units)
  • Log friction points (Effortless Execution tracking)

Effortless Execution Log: Throughout implementation, the skill tracks:

  • Steps that felt heroic (should have been simpler)
  • Friction points encountered
  • Deviations from plan (require explicit approval)
  • Simplification opportunities for future

If friction is systemic, STOP and review the design.

Quality Gate

Uses the quality-gate skill for comprehensive review:

Check When
Code review Always
UBS Scanner (1000+ bug patterns) Always
Requirements traceability Always
Security audit If touching auth, crypto, secrets
Performance check If touching hot paths

Gate: GO/NO-GO decision for verification phase.


4. Right Side of V: Verification and Validation

The right-side-of-v agent orchestrates testing phases after implementation.

flowchart TD
    START([Quality Gate Passed]) --> V1

    subgraph VERIFY["Phase 4: Verification"]
        V1[Run UBS Scanner - always first]
        V2[Build traceability matrix<br/>REQ -> Design -> Code -> Test]
        V3[Execute unit tests]
        V4[Execute integration tests]
        V5[Track coverage metrics]
        V6{Defects found?}
        V1 --> V2 --> V3 --> V4 --> V5 --> V6
    end

    V6 -->|YES| LOOP

    subgraph LOOP["Defect Loop-Back"]
        L1{Defect Type?}
        L2[Requirements gap -> Phase 1]
        L3[Design flaw -> Phase 2]
        L4[Implementation bug -> Phase 3]
        L5[Test gap -> Phase 4]
        L1 --> L2
        L1 --> L3
        L1 --> L4
        L1 --> L5
    end

    V6 -->|NO| VD

    VD{Verification<br/>GO/NO-GO} -->|NO-GO| V1
    VD -->|GO| VAL1

    subgraph VALIDATE["Phase 5: Validation"]
        VAL1[System testing against NFRs]
        VAL2[Conduct UAT interviews]
        VAL3[Verify acceptance criteria]
        VAL4[Document stakeholder feedback]
        VAL5[Gather sign-off]
        VAL1 --> VAL2 --> VAL3 --> VAL4 --> VAL5
    end

    VAL5 --> FINAL{Final<br/>GO/NO-GO}
    FINAL -->|NO-GO| V1
    FINAL -->|GO| RELEASE([Release Ready])

    style VERIFY fill:#e8f5e9
    style VALIDATE fill:#e8f5e9
    style LOOP fill:#ffcdd2
Loading

Phase 4: Verification

Uses the disciplined-verification skill to answer: "Did we build it right?"

Specialist Skills Invoked:

  • ubs-scanner: 1000+ bug patterns across 8 languages
  • requirements-traceability: Matrix construction and gap analysis
  • code-review: Quality verification
  • security-audit: If applicable
  • rust-performance: If applicable
  • testing: Test suite execution

Defect Loop-Back: When defects are found, they're traced to their origin:

Defect Type Returns To
Requirements gap Phase 1 (Research)
Design flaw Phase 2 (Design)
Implementation bug Phase 3 (Implementation)
Test gap Phase 4 (Verification)

Phase 5: Validation

Uses the disciplined-validation skill to answer: "Did we solve the right problem?"

Process:

  1. System testing against non-functional requirements
  2. User acceptance testing (UAT) with stakeholders
  3. Verify acceptance criteria are met
  4. Gather stakeholder sign-off
  5. Final GO/NO-GO for release

5. Handover: Preserving Context

The /handover skill generates a comprehensive handover document for async collaboration:

# Handover Document - [Project Name]

**Date:** [Date]
**Branch:** [Branch]
**Last Commit:** [Hash]

## Progress Summary
### Tasks Completed This Session
1. [Specific task with commit reference]
2. [Another task]

### Current State
**What's Working:** [List]
**What's Blocked:** [List with reasons]

## Technical Context
- Branch, commits, modified files
- Key decisions made
- Open questions

## Next Steps
### Priority 1: [Most critical]
### Priority 2: [Important]
### Priority 3: [Nice to have]
### Priority 4: [Future consideration]

## Related Issues
| Issue | Title | Status |
|-------|-------|--------|
| #123 | [Title] | [Open/Closed] |

Why it matters: AI sessions are ephemeral. Without structured handover, the next session (human or AI) starts from scratch. Handover preserves:

  • What was done
  • Why decisions were made
  • What's next
  • Known issues and blockers

The Complete Flow

flowchart TB
    SESSION["/session-start"] --> LEFT

    subgraph LEFT["LEFT SIDE OF V"]
        direction TB
        L1["Phase 1: Research<br/>(disciplined-research)"]
        L15["Phase 1.5: Quality Eval<br/>(KLS framework)"]
        L2["Phase 2: Design<br/>(disciplined-design)"]
        L25["Phase 2.5: Specification"]
        L26["Phase 2.5: Quality Eval"]
        L1 --> L15 --> L2 --> L25 --> L26
    end

    LEFT --> EXEC

    subgraph EXEC["EXECUTION ORCHESTRATOR"]
        direction TB
        E1["Phase 3: Implementation<br/>(disciplined-implementation)"]
        E15["Phase 3.5: Quality Eval"]
        E2["Quality Gate<br/>(code review, UBS, security)"]
        E1 --> E15 --> E2
    end

    EXEC --> RIGHT

    subgraph RIGHT["RIGHT SIDE OF V"]
        direction TB
        R1["Phase 4: Verification<br/>(disciplined-verification)"]
        R2["Phase 5: Validation<br/>(disciplined-validation)"]
        R1 --> R2
    end

    RIGHT --> RELEASE["Release Ready"]
    RELEASE --> HANDOVER["/handover"]

    style SESSION fill:#b3e5fc
    style LEFT fill:#e3f2fd
    style EXEC fill:#fff3e0
    style RIGHT fill:#e8f5e9
    style RELEASE fill:#c8e6c9
    style HANDOVER fill:#b3e5fc
Loading

Key Principles

The 90% Rule

If it's not a clear YES (90%+ fit), it's a NO. Applied at every gate.

No Heroics

If something feels hard, stop and simplify. Log friction for process improvement.

Explicit Elimination

Document what you decided NOT to do. The "Avoid At All Cost" list is as important as the "To Do" list.

Defect Loop-Back

Never just fix bugs - trace them to their origin. A bug in code might be a gap in research.

Quality Gates Block

No phase transition without passing quality checks. Gates are mandatory, not advisory.


Getting Started

Install terraphim-skills as a Claude Code plugin:

# Add the marketplace
claude plugin marketplace add terraphim/terraphim-skills

# Install the plugin
claude plugin install terraphim-engineering-skills@terraphim-skills

Start a disciplined session:

/session-start

Then invoke the left-side-of-V agent for planning:

Use the left-side-of-v agent to plan implementation of [your feature]

After planning approval, use execution orchestrator:

Use the execution-orchestrator agent to implement the approved plan

After implementation, verify and validate:

Use the right-side-of-v agent to verify and validate the implementation

Finally, create handover:

/handover

Conclusion

Terraphim Skills transforms AI-assisted development from "vibe-based coding" to disciplined engineering. By implementing the V-model as Claude Code skills with quality gates at each phase, it ensures:

  • Research before design: Understand the problem first
  • Design before code: Plan the solution
  • Test at every step: Verify continuously
  • Trace everything: Requirements to tests
  • Preserve knowledge: Structured handover

The framework embodies Essentialism ("less but better") and Effortless ("what if this could be easy?") at every phase. The result: AI that helps you build the right thing, the right way, with full traceability and no heroics.


Resources:


Published: January 2026

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment