Skill Phase Context

---
name: skill-phase-context
description: "Skill loading tiers: TIER 0 (always), TIER 1 (phase-triggered), TIER 2 (extended). Defines when to load which skills."
tier: 2
version: 1.3
---
# Phase Context Loading Protocol

This skill defines **skill loading tiers** to optimize token consumption while preserving automation capabilities.

> [!IMPORTANT]
> **Single Source of Truth for Skill Loading Strategy.**
> Reference this skill to understand WHEN to load WHICH skills.

## TIER 0: System Foundation — ALWAYS LOADED

> [!CAUTION]
> **NEVER lazy-load TIER 0 skills. They enable core automation.**
> These skills MUST be available in EVERY session at bootstrap.

| Skill | Tokens | Why Always Load |
|-------|--------|-----------------|
| `core-principles` | ~519 | Anti-hallucination rules, Stub-First, Documentation First |
| `skill-safe-commands` | ~927 | **Enables automation** — the archive script, `ls`, `git`, tests auto-run |
| `artifact-management` | ~636 | Archiving protocol, file management, dual state tracking |
| **TOTAL** | **~2,082** | **Non-negotiable system foundation** |

### TIER 0 Loading Rule
- Load at session bootstrap (via `GEMINI.md` / `.cursorrules`)
- These skills are loaded BEFORE any phase begins
- Never remove or defer loading of TIER 0 skills

---

## TIER 1: Phase-Triggered — Load on Phase Entry

These skills are phase-specific and loaded when entering the corresponding phase.

| Phase | Skills to Load | Approx. Tokens |
|-------|----------------|----------------|
| **Analysis** | `requirements-analysis`, `skill-task-model`, `artifact-formalizer` (contract only), (`skill-archive-task` if new task) | ~3,100-4,000 |
| **Architecture** | `architecture-design`, `artifact-formalizer` (contract only), (`architecture-format-core` OR `-extended`) | ~2,850-4,200 |
| **Planning** | `planning-decision-tree`, `skill-planning-format`, `tdd-stub-first`, `artifact-formalizer` (contract only) | ~2,525 |
| **Development** | `developer-guidelines`, `documentation-standards` | ~1,100 |
| **Review** | Phase-specific `-review-checklist` skill | ~300-400 |
| **Security** | `security-audit` | ~500 |
| **VDD** | `vdd-adversarial`, (`adversarial-*` per workflow) | ~1,000-2,500 |

### TIER 1 Loading Conditions

**Analysis Phase:**
- `requirements-analysis` — When 02_analyst activated
- `skill-task-model` — When creating/updating TASK.md
- `skill-archive-task` — When existing TASK.md detected AND new (unrelated) task

**Architecture Phase:**
- `architecture-design` — When 04_architect activated
- `architecture-format-core` — **Default** for most updates
- `architecture-format-extended` — Load ONLY for:
  - Creating NEW system from scratch
  - Major refactor (>3 components affected)
  - User explicitly requests "full architecture template"

**Planning Phase:**
- `planning-decision-tree` — When 06_planner activated
- `skill-planning-format` — When creating PLAN.md
- `tdd-stub-first` — When stub creation starts

**Development Phase:**
- `developer-guidelines` — When 08_developer activated
- `documentation-standards` — When updating docs

**Review Phase:**
- Load the appropriate `*-review-checklist` for the current review type

**Every authoring phase (Analysis, Architecture, Planning):**
- `artifact-formalizer` — load `references/authoring-contract.md` **before** the first sentence of
  any artifact is written. Load the SKILL.md body and the scanner only when auditing an artifact
  that already exists.
- **Why.** The contract is ~1,100 tokens; auditing an artifact afterwards costs a full re-read of
  that artifact plus the rewrite. Measured cost of the late order:
  `artifact-formalizer/references/measurement-baseline.md` §5.

---

## TIER 2: Extended — Load Only When Explicitly Requested

These skills are specialized and loaded only when explicitly invoked by user or workflow.

| Skill | Tokens | Loading Condition |
|-------|--------|-------------------|
| `architecture-format-extended` | ~3,000 | New system / major refactor / user requests full template |
| `skill-reverse-engineering` | ~1,322 | User requests "sync docs with code" / docs-code mismatch |
| `skill-update-memory` | ~1,101 | Post-development phase, .AGENTS.md updates needed |
| `skill-adversarial-security` | ~838 | VDD workflow with security focus |
| `skill-adversarial-performance` | ~946 | VDD workflow with performance focus |
| `vdd-sarcastic` | ~145 | VDD workflow with sarcastic mode |

### Condition-loaded: any phase, before the first figure

One TIER 2 skill loads on a condition instead of a request.

- `mermaid-authoring-guidelines` — load `SKILL.md` before the first figure of any output is
  written. Then load the references its routing table names for the figure's kind. It is not
  loaded for an output with no figure.
- A plan chart that `plan_gantt.py` generates needs no load: the script draws it from the schedule
  block (`skill-planning-format` §2.1).
- **Why.** With a load on phase entry, an agent reads the skill in phases that draw no figure. The
  bootstrap files (`CLAUDE.md`, `AGENTS.md`, `GEMINI.md`) state the output-medium rule and its
  exception themselves. A terminal answer follows that rule with no skill loaded.

---

## Integration

### How Agents Should Reference This

Agent prompts should:
1. Always load TIER 0 skills at session start
2. Load TIER 1 skills when entering their specific phase
3. Load TIER 2 skills only when explicitly triggered, or on the condition of a condition-loaded
   skill

### Referenced By
- `.gemini/GEMINI.md` — TIER 0 at bootstrap
- `.cursorrules` — TIER 0 at bootstrap
- All agent prompts — TIER 1 at phase entry
- VDD workflows — TIER 2 when invoked