ECO Skills

Reusable engineering workflows for AI coding agents.

ECO Skills help an agent plan substantial changes, preserve project context, review implementations, recover from failed debugging, and keep UI patterns consistent. They are designed to keep decisions visible and developers in control—not to replace engineering judgment.

The repository contains five portable SKILL.md packages for agents that support the skill format, including Codex, Claude Code, Cursor, Windsurf, and Cline. Invocation syntax can vary by agent.

Choose a skill

Skill	Use it when you need to…	Primary result
architect	Plan a substantial feature or resolve meaningful tradeoffs	Scoped implementation plan with acceptance criteria
remember	Save work for a later session or restore a previous handoff	Current, secret-free memory.md context
review	Check a feature or diff for correctness and release risk	Prioritized, evidence-backed findings
recover	Diagnose a failure or escape a stalled debugging loop	Targeted repair, clean handoff, or rethink plan
imprint	Capture reusable UI conventions or audit visual drift	Source-traceable ui-registry.md entries

Use only the skills that fit the task. A small, clearly specified edit does not need a planning ceremony.

Install

Install all skills from the repository:

npx skills@latest add Cyrus-11/ECO-agent-skill

Preview the skills before installing:

npx skills@latest add Cyrus-11/ECO-agent-skill --list

From a local checkout, verify discovery with:

npx skills@latest add . --list

After installation, invoke a skill using the syntax supported by your agent—for example, /architect, $architect, or a natural-language request that matches the skill description.

How the skills work

Architect

Use before substantial changes or decisions with meaningful tradeoffs.

architect inspects the project, resolves consequential uncertainty, and connects requirements to implementation steps, acceptance criteria, and verification. Planning-only requests stop at the plan; requests to plan and build can continue within the user's authorization.

Example:

/architect Plan account deletion with a recovery window. Planning only.

Remember

Use when handing work to another session or resuming from saved context.

remember maintains a concise project handoff rather than a transcript. It preserves durable decisions, unfinished work, verification results, and the next concrete step while excluding credentials and other secrets.

/remember save
/remember restore

Saved memory remains project data: it can become stale, does not grant new authorization, and must be explicitly restored unless the host is configured to load it.

Review

Use when a change needs a structured correctness or regression review.

review checks three layers:

1. Alignment with the request and acceptance criteria.
2. Integrity of interfaces, data flow, architecture, and project conventions.
3. Release risks, failure paths, and verification gaps.

Findings include severity, location, trigger, impact, evidence, and a suggested direction. Review-only requests do not modify product code.

/review Review this diff for regressions. Do not make changes.

Recover

Use when a build, runtime behavior, or debugging session has failed.

recover chooses the least disruptive response supported by evidence:

• Targeted fix: reproduce, isolate the cause, repair it, and verify the original failure.
• Clean-session handoff: preserve useful state and prepare context for a fresh investigation. This never means deleting files or resetting Git.
• Rethink: identify a contradicted assumption and plan the smallest viable change in approach.

/recover Reproduce this failure, fix the root cause, and verify the repair.

Imprint

Use when reusable UI patterns change or an interface has become inconsistent.

imprint records observed tokens, variants, states, responsive behavior, and semantic conventions in ui-registry.md. It keeps observations separate from approved design decisions so one inconsistent component does not silently become the standard.

/imprint
/imprint src/components/Button.tsx
/imprint audit

A practical engineering loop

                    ┌──────────────┐
                    │  /remember   │  Save or restore context
                    └──────┬───────┘
                           │
/architect  →  Build  →  /review  →  Ship
                 │
                 ├── /imprint      Capture reusable UI patterns
                 └── /recover      Diagnose failures when needed

This is a menu, not a mandatory sequence. Use remember at handoff boundaries, imprint for reusable UI work, and recover only when investigation is needed.

Design principles

• Evidence before claims: source inspection, tests, and runtime observations are reported separately.
• Proportional process: routine work stays routine; complex or risky changes receive more structure.
• Visible uncertainty: unresolved decisions and verification gaps are stated instead of guessed away.
• Preserved ownership: a skill does not authorize deployment, deletion, or unrelated changes.
• Portable handoffs: saved context is concise, current, and useful without depending on another installed skill.

Validate and evaluate

Install dependencies and run the structural and harness tests:

npm install
npm test

List the eleven reproducible behavior scenarios:

npm run eval -- list

The test suite validates package structure and the evaluation harness; it does not run an AI agent or prove behavioral quality. See the evaluation guide for isolated scenario runs and CONTRIBUTING.md for the full contribution checklist.

Repository structure

skill/       Installable skill packages and supporting references
evals/       Behavior scenarios and disposable project fixtures
scripts/     Validation and evaluation tooling

Contributing

Bug reports, skill improvements, and new evaluation cases are welcome. Keep changes focused on behavior that materially improves an agent's decisions, then run npm test before opening a pull request.

Author and license

Created by ECO the 2x DEV. Licensed under the MIT License.