Advanced Context Engineering for Coding Agents (ACE-FCA)

A proven methodology for managing complex codebases and multi-step implementations through disciplined context management, subagent isolation, and progressive compaction.

Core Principle: Frequent Intentional Compaction (FIC)

The Problem: Context quality is the ONLY lever affecting output quality. Each agent interaction is stateless: context in → decision out. Poor context leads to hallucinations, incorrect assumptions, and task failure.

The Solution: Deliberately structure how context flows through development:

1. Selection - Choose only relevant code/info
2. Compression - Condense into compact artifacts
3. Isolation - Shield from distracting noise

Key Insight: Use 40-60% of context window optimally, not 90-100%. More context ≠ better results.

When to Use This Skill

Use when:

• Setting up ACE-FCA workflow for a new or existing project
• Working with large codebases (100k+ LOC)
• Implementing complex, multi-step features
• Context window struggles (hallucinations, irrelevant suggestions)
• Need to document architectural decisions during implementation
• Want disciplined TDD with verification at each step
• Breaking down large tasks into manageable pieces

Skip when:

• Simple, single-file changes
• Greenfield projects with minimal context
• Quick bug fixes that don't require research

The Three-Phase Workflow

Phase 1: Research (Context Isolation)

Purpose: Understand codebase/problem without polluting context.

Process:

1. Launch subagent(s) for file searches and code analysis
2. Subagent investigates, returns summary only (not full files)
3. Subagent dies after task (clean context)
4. Compile findings into compacted research document

Output: docs/research/YYYY-MM-DD-topic-research.md

Template structure:

## Problem Understanding
[2-3 sentences on core issue]

## Relevant Code Locations
- `path/to/file.ts`: [specific responsibility]
- `path/to/other.py`: [specific responsibility]

## Key Dependencies
[Only what matters for this task]

## Constraints & Considerations
[Technical/business constraints]

## Recommended Approach
[High-level direction based on findings]

Why it works: Parent agent receives compressed insights, not overwhelming raw data.

Phase 2: Planning (Executable Specification)

Purpose: Create complete specification before touching code.

Process:

1. Consume research document (not entire codebase)
2. Break work into 2-5 minute tasks
3. Specify exact files, exact changes, exact test expectations
4. Include complete code snippets (not "add validation")
5. Document key architectural decisions as ADRs

Output: docs/plans/YYYY-MM-DD-feature-plan.md

Task specification format:

### Task 3: Add validation to UserService

**File**: `src/services/user_service.py`

**Changes**:
- Add `validate_email()` method
- Call from `create_user()` before DB insert

**Code**:
```python
def validate_email(self, email: str) -> bool:
    pattern = r'^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}$'
    return re.match(pattern, email) is not None

Test expectation: test_create_user_invalid_email() should raise ValidationError

Verification: pytest tests/test_user_service.py -k email

**ADR Integration**: For significant architectural choices made during planning, create ADR before implementation:
```markdown
### Task 2: Document database choice (ADR)

**Create**: `docs/adr/0003-use-postgresql-for-user-data.md`

**Reason**: Multi-tenancy requires JSONB support and row-level security

**Alternatives considered**: MySQL, MongoDB

**Decision**: PostgreSQL for ACID + JSONB

**Template**: Use `templates/adr-template.md`

Why it works: Plan becomes compacted context for implementation. Tasks are unambiguous. ADRs capture "why" for future maintainers.

Phase 3: Implementation (TDD + Progressive Compaction)

Purpose: Execute plan step-by-step with verification.

Process:

1. Select one task from plan
2. Write test first (if not present)
3. Watch test fail (RED) - don't skip!
4. Write minimal code to pass (GREEN)
5. Refactor if needed
6. Run full test suite
7. Commit with clear message
8. After every 3-5 tasks: Update plan with progress to prevent context drift

Progressive compaction example:

✅ Task 1: Add User model - COMPLETE
✅ Task 2: Document DB choice (ADR) - COMPLETE
✅ Task 3: Add validation - COMPLETE
⏳ Task 4: Email uniqueness check - IN PROGRESS
   - Test written and failing
   - Working on implementation
❌ Task 5: Add API endpoint - TODO
❌ Task 6: Integration test - TODO

Why it works: Plan stays current without re-reading completed work. Context remains in 40-60% optimal range.

Architectural Decision Records (ADRs)

Integration with ACE-FCA: Document significant decisions during Planning phase or Implementation phase when they emerge.

When to create ADR:

• Technology/framework choices
• Database schema decisions
• API design patterns
• Security model choices
• Performance trade-offs
• Architectural patterns

ADR Workflow:

1. During planning, identify decision points
2. Add ADR task to plan before implementation task
3. Create ADR using template (see templates/adr-template.md)
4. Reference ADR in related implementation tasks
5. Commit ADR before implementing decision

File naming: docs/adr/NNNN-title-with-hyphens.md (e.g., 0001-use-react-for-frontend.md)

Benefits:

• Future maintainers understand "why"
• Prevents re-litigating decisions
• Documents alternatives considered
• Creates knowledge base

Subagent-Driven Development

When to use: Implementation plan with mostly independent tasks.

Process:

1. Dispatch fresh subagent for each task
2. Subagent implements following TDD
3. Two-stage review:
   • Stage 1: Spec compliance check
   • Stage 2: Code quality review
4. If review fails, subagent fixes
5. Mark complete, move to next

Why it works: Fresh context per task prevents pollution. Separation of implementation and review maintains objectivity.

Context Quality Framework

Target: 40-60% of context window usage

Too Little (<40%):

• Missing critical dependencies
• Reinventing existing solutions
• Breaking existing patterns → Fix: Include more relevant files in research

Too Much (>60%):

• Model struggles to identify relevant info
• Increased hallucinations
• Slower responses → Fix: Use subagents for isolation, create more compressed artifacts

Measurement: Monitor which files/docs agent references. If citing everything, context too broad. If missing obvious connections, too narrow.

Critical Principles

YAGNI (You Aren't Gonna Need It)

• Remove features from designs
• Implement only what's needed now
• Resist "might need later"
• Simpler designs complete faster

TDD Non-Negotiable

• Tests first, always
• Watch test fail (RED) - confirms test works
• Minimal code to pass (GREEN)
• Refactor for quality
• Tests are safety net for refactoring

Incremental Validation

• Present work in chunks
• Get feedback early and often
• Don't surprise with huge deliverables
• Course-correct early

Spec-First Development

• Write spec before code
• Get human validation on spec
• Implement to spec
• Review against spec → Most bugs are spec bugs, not code bugs

Project Setup for ACE-FCA

When onboarding a new project:

1. Create directory structure:

mkdir -p docs/{research,plans,adr}

2. Initialize ADR log:

# Create docs/adr/README.md with:
# - ADR naming convention
# - List of all ADRs with links
# - Status of each ADR (accepted, superseded, deprecated)

3. Document current state (if existing codebase):

   • Launch research subagent
   • Map key files and responsibilities
   • Identify testing patterns
   • Note architectural constraints
   • Create docs/research/YYYY-MM-DD-codebase-overview.md

4. Set up TDD environment:

   • Verify test runner works
   • Confirm test watching (if applicable)
   • Document test commands in README

5. First ADR: Document decision to use ACE-FCA

   • docs/adr/0001-adopt-ace-fca-methodology.md
   • Explains why, alternatives considered, consequences

Common Patterns

Pattern: Breaking Down Large Feature

1. Research phase → Understand existing architecture
2. Create high-level plan → Break into 5-10 major chunks
3. For each chunk:
   a. Detailed task breakdown (2-5 min tasks)
   b. Identify ADR needs
   c. Implement with TDD
   d. Update progress in plan
4. Final integration testing

Pattern: Debugging with ACE-FCA

1. Research phase → Reproduce bug, understand context
2. Planning phase → Identify root cause, design fix
3. Implementation:
   a. Write test that fails with bug
   b. Fix bug (test goes green)
   c. Add regression test if needed
   d. Document in commit message or ADR if architectural

Pattern: Refactoring Safely

1. Research → Map dependencies and usage
2. Plan → Define transformation steps with test strategy
3. Implement:
   a. Ensure 100% test coverage of code to refactor
   b. Make smallest possible change
   c. Run tests (should stay green)
   d. Commit
   e. Repeat for next small change

Anti-Patterns to Avoid

❌ Jumping straight to code - No research or planning phase
→ Results in rework, bugs, frustration

❌ Vague plans - "Add validation" without specifics
→ Agent guesses, guesses wrong

❌ Loading full codebase - "Here's the entire repo"
→ Agent overwhelmed, generic suggestions

❌ Long tasks without checkpoints - 30+ minute tasks
→ Context drift, agent loses thread

❌ Skipping RED in TDD - Writing test and code together
→ Test might not be testing anything

❌ No ADRs - Decisions made without documentation
→ Future confusion, re-litigation of settled decisions

References

For detailed information, see:

• Context Management Deep Dive - Context window optimization techniques
• Patterns & Anti-Patterns - Proven approaches and what to avoid
• Subagent Workflows - Advanced subagent patterns and coordination
• GitHub Copilot Customization - Integration with agents, prompts, and instructions

Templates

Available in templates/ directory:

Core Workflow:

• research-template.md - Research phase output structure
• plan-template.md - Implementation plan structure
• adr-template.md - Architectural Decision Record format

GitHub Copilot Integration (templates/copilot/):

• ace-fca-coordinator.agent.md - Main coordinator agent
• prompts/research-phase.prompt.md - Quick research start
• prompts/planning-phase.prompt.md - Quick planning start
• prompts/implementation-phase.prompt.md - Quick implementation start
• prompts/create-adr.prompt.md - Quick ADR creation
• ace-fca-workflow.instructions.md - Core workflow rules for copilot-instructions.md

Success Metrics

From ACE-FCA case study (300k LOC Rust codebase):

• Week's worth of work completed in a day
• Code quality passed expert review
• Developer was amateur in that language
• Success: context engineering, not model intelligence

Key enablers:

• Research phase isolated complexity
• Plan provided clear roadmap
• TDD discipline caught regressions
• Progressive compaction maintained context quality
• ADRs documented decisions for future

Quick Reference Card

Context flow: Problem (large) → Research Doc (compact) → Plan (executable) → Code (tested)

Golden rule: If context feels overwhelming, compact more aggressively. The answer is always better compression, not bigger context windows.