Using with AI Assistants

codeknit ships with ready-made skills that teach AI coding assistants how to use it effectively. These skills enable assistants to extract code structure, detect duplicates, and perform structural analysis without manual prompting.

Skills overview

codeknit provides two skills:

• codeknit-parse: Teaches assistants to extract code structure (functions, classes, methods, variables) and relationships (calls, inheritance, containment) into .skt files.
• codeknit-fingerprint: Teaches assistants to detect duplicate and near-duplicate code using fuzzy hashing.

Each skill includes documentation that the assistant reads on demand to understand usage, flags, output formats, and workflows.

Installation

Use the installer helper to copy the skill directories to your assistant’s skills folder. The installer downloads only the bundled skill files, so you do not need to clone the repository.

Install for Codex, Kiro, and Claude Code:

curl -fsSL https://raw.githubusercontent.com/awslabs/codeknit/main/scripts/install-skills.sh | bash

Install for one assistant:

curl -fsSL https://raw.githubusercontent.com/awslabs/codeknit/main/scripts/install-skills.sh | bash -s -- --assistant codex
curl -fsSL https://raw.githubusercontent.com/awslabs/codeknit/main/scripts/install-skills.sh | bash -s -- --assistant kiro
curl -fsSL https://raw.githubusercontent.com/awslabs/codeknit/main/scripts/install-skills.sh | bash -s -- --assistant claude

From a local checkout, you can use Makefile helpers:

make skills-install-dry-run
make skills-install

The installer skips existing skill directories by default. To replace them, add --force:

curl -fsSL https://raw.githubusercontent.com/awslabs/codeknit/main/scripts/install-skills.sh | bash -s -- --assistant all --force

After installation, the assistant automatically knows how to invoke codeknit commands, select appropriate flags, and interpret .skt output.

What each skill teaches

codeknit-parse

The codeknit-parse skill teaches assistants to:

• Run codeknit parse with appropriate flags for different scenarios
• Choose the right output mode:
  • directory-flat (default) for most projects
  • inline for single files or small inputs
  • directory-tree to mirror source structure
• Read and interpret .skt output files, including [symbols], [edges], and optional [dict] sections
• Use structural data for refactoring, dependency mapping, and code review
• Run codeknit graph analyze for deeper code quality insights (cyclic dependencies, hub symbols, god classes, etc.)

codeknit-fingerprint

The codeknit-fingerprint skill teaches assistants to:

• Use codeknit fingerprint for duplicate detection, DRY audits, and refactor identification
• Select appropriate similarity ranges (--min-similarity, --max-similarity)
• Read the [duplicates] section to identify near-duplicate code
• Understand that fingerprints measure structural shape, not semantic intent
• Use --rerank with Ollama embeddings to reduce false positives when needed

Workflow examples

Structural analysis

1. Ask the assistant to analyze your codebase structure
2. It runs codeknit parse ./src and reads the resulting .skt files
3. It answers structural questions: dependencies, call chains, dead code
4. For deeper insights, it runs codeknit graph analyze ./src and interprets the report

[symbols]
## src/service.go
S1 type/struct L5-L8 AuthService {}
S2 callable/method L10-L15 Authenticate(token: string) {receiver=*AuthService}

[edges]
S1 --contains--> S2

Duplicate detection

1. Ask the assistant to find duplicated code
2. It runs codeknit fingerprint ./src
3. It reads the [duplicates] section in the output
4. It investigates flagged pairs and proposes consolidation

[duplicates]
S1, S2: 87% similarity
S3, S4: 76% similarity

Tips

• Always read .skt files, not raw source, for structural questions — they contain the extracted structure in a compact, reliable format
• Use codeknit graph analyze to uncover code quality issues like cyclic dependencies, hub symbols, and deep inheritance chains
• Run codeknit fingerprint before large refactors to identify copy-pasted code that should be consolidated
• The .skt format is designed to be token-efficient, making it ideal for LLM context windows
• Use --minify to further reduce token usage when processing large codebases