Convention Extraction

Most team conventions live in people's heads — pattern recognition built from years of reviews, production incidents, and architectural discussions. They transfer slowly through pairing and code review, and walk out the door when someone leaves. AI amplifies this: without explicit conventions, AI output quality varies by who prompts. Same codebase, same AI, completely different quality gates.

This skill guides systematic extraction of tacit knowledge into versioned, enforceable artefacts. The approach is informed by Rahul Garg's "Encoding Team Standards" (2026), which frames inconsistent AI output as a systems problem requiring a systems solution.

This skill does not cover convention enforcement (see constraint-design and verification-slots), convention maintenance (see context-engineering and garbage-collection), or CI pipeline configuration.

For the full interview protocol with worked examples, consult references/extraction-interview-guide.md.

When to Extract

Situation	Signal
New project setup	CLAUDE.md and HARNESS.md are empty or boilerplate
Onboarding AI to existing codebase	AI-generated code keeps violating unwritten rules
After team composition changes	New members or departures change the tacit knowledge base
Quality variance	AI output quality depends on who is prompting
Post-incident	A production incident reveals conventions that were implicit

Sizing heuristic: Teams of five may not need formal extraction — conversations happen naturally. Teams of fifteen almost certainly do — tacit knowledge diverges without intervention.

The Five Extraction Questions

These questions surface the tacit knowledge that matters most for AI collaboration. Ask them in a team setting (mob session) or through structured one-on-one interviews with senior engineers.

1. What architectural decisions should never be left to individual judgment? Surfaces non-negotiable patterns — dependency direction, module boundaries, API design rules.

2. Which conventions are corrected most often in AI-generated code? Surfaces the gap between what AI produces by default and what the team expects. These are your highest-value conventions.

3. Which security checks are applied instinctively? Surfaces embodied security knowledge — input validation, auth patterns, secrets handling — that seniors apply without thinking.

4. What triggers an immediate rejection in review? Surfaces hard boundaries — the things that are never acceptable regardless of context.

5. What separates a clean refactoring from an over-engineered one? Surfaces judgment about abstraction thresholds, YAGNI boundaries, and when to stop.

Mapping Answers to Artefacts

Each answer maps to a specific artefact type and location:

Answer category	Priority tier	Artefact type	Where it lives
Non-negotiable patterns	Must-follow	Constraint	HARNESS.md
Frequent corrections	Should-follow	Convention	CLAUDE.md
Security instincts	Must-follow	Threat-model item	HARNESS.md or security skill
Review rejections	Must-follow	Critical check	Code reviewer agent instructions
Refactoring philosophy	Nice-to-have	Style preference	CLAUDE.md or skill reference
"It depends" answers	Not encodable yet	Aspiration	Backlog for decomposition

"It depends" is a signal, not a failure. If the answer is always "it depends on context," the convention needs decomposition into specific, observable cases before it can be encoded.

The Four-Element Anatomy

Every encoded instruction benefits from four elements:

1. Role Definition

What expertise is assumed? Not persona play — it establishes the lens. "A senior engineer implementing a new service following the team's architectural patterns" is different from "a security reviewer assessing input handling."

2. Context Requirements

What does the instruction need to operate? Relevant code, architectural context, applicable constraints. Makes dependencies explicit.

3. Categorised Standards

Organised by priority tier:

Tier	Meaning	Enforcement
Must-follow	Non-negotiable. Violations are blockers.	HARNESS.md constraint
Should-follow	Strong expectation. Exceptions need justification.	CLAUDE.md convention
Nice-to-have	Preferred but not enforced.	CLAUDE.md or skill reference

4. Output Format

What should the result look like? A structured response with summary, categorised findings, and clear next steps. Ensures consistency across developers and sessions.

Anti-Patterns

Anti-pattern	Problem	Fix
Over-prescriptive instructions	Brittle, false positives on edge cases	Test the instruction against real code before committing
Encoding aspirations	"Write clean code" is not enforceable	Decompose into observable properties
Documentation graveyards	Created with enthusiasm, abandoned in months	Place artefacts close to the workflow; use GC rules for freshness
Premature scaling	Ten instructions before one is adopted	Start with one instruction (generation or review) and expand after adoption
Skipping disagreement	Assuming seniors agree when they don't	The extraction conversation is the point — disagreements are features

When to Re-Extract

• After significant team composition changes (someone joins or leaves)
• After a production incident reveals tacit knowledge that wasn't encoded
• When AI output quality degrades without obvious cause
• Quarterly, as part of the operating cadence — review whether encoded conventions still match practice

The Command

Run /extract-conventions for a guided, conversational extraction session that walks through the five questions and maps answers to artefacts.