Enforce mandatory context loading before AI code changes

the-honoured-one - Full Context Load Protocol

Overview

Gojo at full power means all six eyes open - everything visible, nothing assumed, no blind spots. The Honoured One doesn't act on guesses. This skill enforces the same: the AI must earn the right to act by reading and understanding first.

The most common AI coding failure is confident wrongness - the AI proposes or implements something based on how it assumes the code is structured, not how it actually is. It gets the architecture wrong, uses a pattern inconsistent with the rest of the codebase, or integrates with a module it never actually opened. This skill eliminates that failure mode by making context-loading mandatory before any action.

When to Use This Skill

• Use when modifying multiple files in an existing codebase
• Use when designing or modifying a system component
• Use when adding a feature that integrates with existing code
• Use when debugging a system or component the AI has not yet read
• Use when the AI would need to assume how existing code is structured
• DO NOT use for isolated single-file tasks where the file has already been read

How It Works

PHASE 1 - Context Audit

When given any complex task, the AI must immediately perform a context audit before proposing anything. It declares:

1. What files are relevant to this task? - Every file that will be read, changed, or is upstream/downstream of the change
2. Which of those has the AI actually read this session? - Honest accounting, no assumptions
3. What gaps exist? - Files that are relevant but unread

The AI outputs this before doing anything else:

THE HONOURED ONE — CONTEXT AUDIT
─────────────────────────────────────────
Task: [what was asked]

Relevant files identified:
  - src/auth/middleware.ts       → [why relevant]
  - src/routes/user.ts          → [why relevant]
  - src/models/user.model.ts    → [why relevant]
  - src/utils/token.ts          → [why relevant]

Files read this session:
  - src/routes/user.ts          → ✓ read

Unread but relevant (blind spots):
  - src/auth/middleware.ts       → ✗ not read
  - src/models/user.model.ts    → ✗ not read
  - src/utils/token.ts          → ✗ not read
─────────────────────────────────────────
Cannot proceed — reading blind spots now.

The AI cannot propose a solution, make a plan, or write any code while blind spots exist.

PHASE 2 - Mandatory Read Pass

The AI reads every file listed as a blind spot. Not summaries, not assumptions based on filename or folder structure - actual reads.

Rules for this phase:

• If a file imports from another file that is also relevant, that file gets added to the read list
• If reading a file reveals unexpected structure or patterns, the AI notes this before continuing
• The AI does not form opinions or solutions while reading - this phase is observation only

Shortcut rule: The AI cannot say "I'm familiar with this pattern so I don't need to read it." Familiarity with a pattern is not familiarity with this codebase's implementation of it.

PHASE 3 - Orientation Statement

After all relevant files are read, the AI outputs an orientation statement before proposing anything. This is its proof that it understands the codebase well enough to act:

THE HONOURED ONE — CONTEXT LOADED
─────────────────────────────────────────
Files read: [complete list]

Current architecture (what I now know):
  [2-3 sentences describing how the relevant system actually works,
   based on what was read — not assumed]

What this task touches:
  - [file/component 1] → [how it's involved]
  - [file/component 2] → [how it's involved]

Existing patterns I must follow:
  - [naming convention / error handling style / structure pattern observed]
  - [any other conventions seen in the actual code]

Remaining unknowns:
  - [anything still unclear — or "None, ready to proceed"]
─────────────────────────────────────────

PHASE 4 - Confidence Gate

After the orientation statement, the AI applies a confidence gate before acting:

If "Remaining unknowns" is empty:
→ Proceed. The AI is fully loaded and may propose a solution or begin work.

If "Remaining unknowns" is non-empty:
→ The AI must resolve every unknown before proceeding. Options:

• Ask the user the specific question
• Read another file that would answer it
• Acknowledge the unknown, state the assumption being made, and get user confirmation before continuing

The AI cannot proceed with known blind spots. Stating "I'll assume X" and moving forward without user confirmation is not allowed.

Self-Ask Before Acting

Before writing any code or making any proposal, the AI must answer:

Hard Rules (Never Violated)

• No proposing solutions before reading. Proposals based on assumptions are not proposals - they are guesses.
• No "I assume this file does X." If you haven't read it, you don't know what it does.
• No skipping files because their names look obvious. A file called utils.ts can contain anything.
• No importing or calling code from files that haven't been read. You cannot use what you haven't seen.
• No "familiar pattern" shortcuts. The pattern may be implemented differently here.
• No acting with known unknowns. Resolve them or get user confirmation before proceeding.

What This Skill Prevents

• AI proposing integration with a module structured completely differently than assumed
• AI using naming conventions inconsistent with the rest of the codebase
• AI calling functions that don't exist because it assumed they would be there
• AI making architectural decisions that conflict with patterns already established in the code
• AI confidently implementing the wrong thing and needing a full redo

Quick Reference

Trigger Phrases

• "add this feature to the existing code"
• "integrate X with Y"
• "modify how [system] works"
• "refactor this"
• "why is this not working" (on unread code)
• Any task touching more than one file
• Any task where the AI would need to know how existing code is structured to do it correctly

Examples

(Examples of the Context Audit and Orientation Statement output are provided inline within Phase 1 and Phase 3 above.)

Best Practices

• ✅ Do: Ensure all blind spots are read before proceeding.
• ✅ Do: Confirm the AI's orientation statement matches reality.
• ❌ Don't: Allow the AI to skip reading just because a file name seems obvious.

Common Pitfalls

• Problem: The AI assumes an implementation matches a common pattern without reading it.
  Solution: Enforce Phase 2 (Mandatory Read Pass) without exceptions.

Related Skills

• @brainstorming - Use before execution to figure out what needs to be built.
• @not-a-vibe-coder - Use for entirely new projects, whereas this skill is for existing ones.

Limitations

• This skill requires more token usage due to reading multiple files upfront.
• It may slow down the initial response time before the AI starts coding.
• The AI might end up reading more files than strictly necessary if the dependency chain is deep.
• Does not replace the need for the user to verify the final code.