Coherence

Five portable Agent Skills for Claude Code that keep humans and coding agents aligned at high velocity. Each skill is a single directory containing one SKILL.md with no scripts and no dependencies — drop it into your agent's skills directory and it works.

Why

These skills came out of building Elisa, an IDE for kids learning agentic programming, during the Built with Opus 4.6 hackathon. The project moved faster than any human could manually inspect every TypeScript change. The lesson: agentic coding only succeeds when the human and the agent maintain context coherence — shared intent, explicit boundaries, verified behavior, and disciplined resets when context drifts.

These are the operating procedures that fall out of that lesson. Not motivational prompts. Concrete, checklist-driven workflows.

The skills

Skill	Use when
boundary-definition	Starting a feature, refactor, integration, API change, or state-machine change. Forces explicit interfaces, contracts, invariants, ownership, non-goals, and failure modes before code.
multi-pass-planning	Task is non-trivial, spans multiple files, or has unclear assumptions. Runs Plan -> Audit -> Repair -> Align -> Code as five distinct passes.
verification-as-context	Implementing behavior, choosing test level, or deciding what proof must ride with a change. Treats tests, docs, and CI as the executable record of product intent.
manual-compaction-reset	Context is long, a session resumes after time away, or the agent is carrying stale assumptions. Replaces fuzzy chat memory with a curated, written brief.
collaboration-checkpoint	Repeated fixes fail, progress stalls, or the user says the agent is missing something. Pauses code churn and surfaces the user's mental model.

How they compose

boundary-definition  ->  multi-pass-planning  ->  code
                                 |
                                 +-->  verification-as-context  (during planning and before completion)

context drift   ->  manual-compaction-reset
debug stall     ->  collaboration-checkpoint

Each skill stands alone. The compositions are recommendations, not requirements.

Install

Copy any skill into your agent's skills directory:

# Claude Code (user-scoped)
cp -r skills/boundary-definition ~/.claude/skills/

# Or symlink the whole set
ln -s "$(pwd)/skills" ~/.claude/skills

Verify the agent can see them — in Claude Code, the skill names should appear in the available-skills list at session start.

Format

Each SKILL.md follows the same structure:

• YAML frontmatter (name, description)
• When to use this skill
• Core principle
• Procedure
• Required output format
• Stop conditions / escalation triggers
• Short examples

Directory name matches the YAML name exactly. Lowercase, hyphens only.

License

MIT. See LICENSE.