Evidence-driven software engineering workflows for coding agents.
A practical catalog for Codex, Claude Code, Cursor, OpenCode, and other Agent Skills-compatible tools. The skills are designed to make agents slow down at the decisions that matter, work in verifiable increments, and distinguish a plausible implementation from a proven one.
Governing rule: an agent may propose freely, but it may only claim success after exercising the relevant system boundary and recording the result.
Spanish documentation · Catalog · Contributing · Security
Coding agents often fail in predictable ways:
- they start editing before understanding the requested behavior;
- they plan horizontal layers instead of complete user-visible slices;
- they anchor on the first plausible debugging theory;
- they treat passing unit tests as proof that the system works;
- they report completion without showing reproducible evidence;
- they mix review concerns until important risks disappear inside generic feedback.
This catalog packages software-engineering discipline into small, composable skills with explicit completion criteria and reusable artifacts.
CLARIFY → SPECIFY → SLICE → IMPLEMENT → DEBUG / REVIEW → VERIFY → SHIP
Each stage produces a concrete artifact:
| Stage | Primary artifact | Question answered |
|---|---|---|
| Clarify | Decision Record | What behavior and constraints are actually intended? |
| Specify | Change Specification | What must be true when the change is complete? |
| Slice | Delivery Plan | What is the smallest reversible increment that proves progress? |
| Implement | Evidence Packet | What changed, and what targeted check reacted? |
| Debug | Evidence Ledger | Which observations support or contradict each hypothesis? |
| Review | Four-Lens Review | Does the change satisfy intent, behavior, risk, and design? |
| Verify | Runtime Proof | Does the complete system path work? |
| Ship | Release Evidence | What supports release, monitoring, and rollback? |
npx skills@latest add FernandoAbishai/agent-skillsList the available skills without installing:
npx skills@latest add FernandoAbishai/agent-skills --listInstall the complete catalog globally for Codex and Claude Code:
npx skills@latest add FernandoAbishai/agent-skills \
--skill '*' \
--agent codex \
--agent claude-code \
--global \
--yesStandalone invocation:
Claude Code: /fernandoabishai <task>
Codex: $fernandoabishai <task>
/plugin marketplace add FernandoAbishai/agent-skills
/plugin install fa-engineering@fernandoabishai
Plugin invocation is namespaced:
/fa-engineering:fernandoabishai <task>
| Skill | Purpose | Main output |
|---|---|---|
fernandoabishai |
Route work to the smallest trustworthy workflow | Route card |
setup-engineering-context |
Discover repository commands, boundaries, and architecture facts | Engineering context |
clarify-change |
Resolve load-bearing product and safety decisions | Decision record |
write-change-spec |
Define behavior without prescribing implementation | Change specification |
plan-delivery |
Break work into vertical, reversible increments | Slice plan |
implement-change |
Implement one verified slice at a time | Evidence packet |
| Skill | Purpose | Main output |
|---|---|---|
debug-with-evidence |
Diagnose failures with competing hypotheses and discriminating experiments | Evidence ledger |
review-change |
Review intent, behavior, risk, and design independently | Four-lens review |
verify-system |
Prove the complete runtime story through real boundaries | Verification matrix |
ship-with-evidence |
Prepare a release decision with monitoring and rollback | Release evidence report |
See the engineering catalog for invocation guidance, composition rules, and examples.
/fernandoabishai Add team invitations with expiring links
Expected route:
clarify-change → write-change-spec → plan-delivery → implement-change
→ review-change → verify-system → ship-with-evidence
/fernandoabishai Users occasionally receive another user's cached dashboard
Expected route:
debug-with-evidence → implement-change → review-change → verify-system
/fernandoabishai Review this branch against issue #184
Expected route:
review-change → verify-system when runtime evidence is missing
Every promoted skill follows four rules:
- One recognizable failure mode. A skill must correct a specific agent behavior.
- Checkable completion. The agent must be able to distinguish done from merely plausible.
- Progressive disclosure. Core steps remain in
SKILL.md; templates and detailed references live beside the skill and load only when needed. - Action safety. Skills do not authorize pushing, merging, deploying, modifying production data, rotating secrets, or destructive migrations.
Each skill directory may contain:
skill-name/
├── SKILL.md
├── agents/openai.yaml
├── references/
├── templates/
└── examples/
Resources are co-located with the skill so installing one skill does not leave broken references.
A skill is not considered improved because its prose sounds stricter. It must create an observable behavioral difference.
Each evaluation should compare:
same repository + same task + same agent
without skill vs. with skill
Useful pass criteria include:
- fewer unsupported completion claims;
- earlier discovery of missing requirements;
- a reproducible bug signal before repair;
- smaller and more reversible implementation increments;
- review findings tied to concrete failure scenarios;
- runtime verification at the boundary named in the specification.
Evaluation fixtures and documented results will be added incrementally rather than fabricated.
This is an early engineering release. The initial lifecycle is usable, but the catalog is not presented as universally superior to established repositories. Its differentiator is narrower: evidence, completion criteria, runtime proof, and action safety are first-class concerns.
Planned additions:
- change-impact analysis;
- safe database and data migrations;
- dependency upgrades;
- architecture and boundary design;
- performance investigation;
- merge-conflict resolution;
- issue triage and delivery tracking;
- behavioral evaluation fixtures.
This repository uses the open Agent Skills format and is informed by established software-engineering practices and the broader public skills ecosystem. The workflows, terminology, artifacts, completion criteria, and supporting resources in this repository are original unless a file explicitly states otherwise.
It is not a translation or a renamed mirror of another catalog.
MIT © 2026 Fernando Abishai. See LICENSE.