# Dev Workflows
**Agent workflows as `SKILL.md` files for frontend, backend, architecture, security review and READMEs.
Every phase ends in a verification step, not a claim.**
[Workflows](#workflows) · [Install](#install) · [How it works](#how-it-works) · [QA harness](#the-qa-harness) · [Known limits](#known-limits)
**[Ler em português](README.pt-BR.md)**
> [!NOTE]
> **Project state:** five commits between 2026-08-19 (workflows, bilingual README,
> `wf-readme`) and 2026-08-24 (ten security skills vendored under `skills/`). Every
> workflow declares `version: 1.0.0`. The workflow bodies are written in Portuguese;
> this README is in English and [README.pt-BR.md](README.pt-BR.md) is the Portuguese copy.
> [!IMPORTANT]
> **Only `wf-security-review` works out of the box.** Its skills are vendored in
> [`skills/`](skills/NOTICE.md). The frontend, backend, architecture and README
> workflows load third-party skills that are **not** in this repository and must be
> installed separately ([list below](#skills-the-workflows-load)). The security
> workflow also needs network access: it must query OSV.dev, the GitHub Advisory DB
> and NVD, and stamp the date on the report.
## Workflows
| Workflow | File | Phases | Covers |
|---|---|---|---|
| **Frontend / UI** | [`wf-frontend.md`](workflows/wf-frontend.md) | DESIGN, BUILD, REVIEW | Design direction before code, anti-AI-slop, React performance, visual QA for PDFs |
| **Backend / API** | [`wf-backend.md`](workflows/wf-backend.md) | PLAN, BUILD, SECURITY, TEST + BUG HUNT | Contract before handler, safe defaults, OWASP audit, tests and adversarial bug hunt |
| **Architecture** | [`wf-architecture.md`](workflows/wf-architecture.md) | REQUIREMENTS, DESIGN, DOCS, REVIEW | Measurable requirements, C4 diagrams, ADRs, threat modeling |
| **Security review** | [`wf-security-review.md`](workflows/wf-security-review.md) | SCOPE, HUNT, CONSULT, CODE REVIEW, REPORT + GATE | Front and back code review, known CVEs and a live lookup of new ones |
| **README** | [`wf-readme.md`](workflows/wf-readme.md) | SCAN, CLASSIFY, AUDIT or DRAFT, GENERATE, VERIFY + VALIDATE | Create, audit or rewrite READMEs grounded in real repo facts |
Each file is a standard `SKILL.md`: the `description` in the frontmatter is the
trigger, so the workflow loads when the agent's context matches it.
Why they exist: agents generate code well and admit ignorance badly. Each workflow
forces the agent to prove the work at every phase:
- **Verification steps:** run the build, run the tests, hit the API, consult the registry.
- **Acceptance criteria:** checkbox gates that must be marked before a phase closes.
- **Hard rules:** "never merge with a red build", "never claim 'no CVE' without querying OSV.dev".
- **A QA harness:** a scorecard for workflow quality, plus meta-tests that corrupt the harness to prove it catches regressions.
## Install
```bash
# 1. Copy the workflows into your agent's skills directory
# Hermes Agent
cp workflows/*.md ~/.hermes/skills/software-development/
# Claude Code: each skill needs its own directory (a loose .md is not loaded)
mkdir -p ~/.claude/skills
for f in workflows/*.md; do
name=$(basename "$f" .md)
mkdir -p "$HOME/.claude/skills/$name"
cp "$f" "$HOME/.claude/skills/$name/SKILL.md"
done
# 2. Security review: the supporting skills are already vendored in skills/.
# Hermes:
cp -r skills/* ~/.hermes/skills/
# Claude Code: copy each skills// directory to ~/.claude/skills//
# (the SKILL.md must sit inside its own directory, as in step 1).
# 3. Use it
# "review this PR for security" -> wf-security-review loads
# "build a new page" -> wf-frontend loads
```
The remaining skills install through `npx skills` (agentskills.io format), for example:
```bash
npx skills add anthropics/skills --skill frontend-design
npx skills add vercel-labs/agent-skills --skill vercel-react-best-practices
npx skills add conorbronsdon/avoid-ai-writing
npx skills add https://github.com/pbakaus/impeccable --skill impeccable
```
> [!WARNING]
> **Install layout matters.** Claude Code ignores a loose `.md`; it needs
> `/SKILL.md`. The QA harness has the same requirement: it reads
> `~/.hermes/skills/software-development//SKILL.md`, so the flat `cp`
> used for Hermes above is not enough to run it (see [the harness](#the-qa-harness)).
## How it works
The security workflow is built on one fact: a model's knowledge of CVEs stops at
its training date, so "no vulnerabilities" from memory is a guess.
```mermaid
flowchart LR
A["Code / diff"] --> B["1 SCOPE"]
B --> C["2 HUNT
known CVEs"]
C --> D["3 CONSULT
OSV.dev, GitHub Advisory, NVD"]
D --> E["4 CODE REVIEW
OWASP Top 10 + ASVS"]
E --> F["5 REPORT + GATE
sources and date stamped"]
```
A review with no source and no date is expired on arrival.
### Skills the workflows load
Vendored in skills/ (security review)
Exact origin, commit and license of each copy are in [`skills/NOTICE.md`](skills/NOTICE.md).
The copies are 1:1 with the origin; CC-BY-4.0 requires attribution, kept in that file.
| Skill | What it does | Origin | License |
|---|---|---|---|
| `safedeps` | Queries OSV.dev live: CVE, severity and fixed version for a `package@version` (script `check_deps.py`) | [Jeneidi/safedeps](https://github.com/Jeneidi/safedeps) | MIT |
| `sca-audit` | Dependency supply-chain audit with reachability analysis and CWE mapping | [OWASP/secure-agent-playbook](https://github.com/OWASP/secure-agent-playbook) | CC-BY-4.0 |
| `code-review-security` | Systematic security code review mapped to OWASP Top 10 + ASVS | OWASP/secure-agent-playbook | CC-BY-4.0 |
| `secrets-scan` | Hardcoded credentials and API keys in code and git history | OWASP/secure-agent-playbook | CC-BY-4.0 |
| `api-security-review` | API review against OWASP API Security Top 10 | OWASP/secure-agent-playbook | CC-BY-4.0 |
| `web-security-review` | Web app review against OWASP Top 10 | OWASP/secure-agent-playbook | CC-BY-4.0 |
| `cve-triage` | Prioritizes CVEs with CVSS 4.0, EPSS and CISA KEV | [UnitOneAI/SecuritySkills](https://github.com/UnitOneAI/SecuritySkills) | MIT |
| `patch-prioritization` | Decides remediation order | UnitOneAI/SecuritySkills | MIT |
| `dependency-scanning` | Dependency tree scanning | UnitOneAI/SecuritySkills | MIT |
| `hermaguard` | Adversarial review: static pre-scan plus 3 specialist agents (edge cases, attack, blast radius) | [Sahil-SS9/hermaguard](https://github.com/Sahil-SS9/hermaguard) | MIT |
Not vendored (install separately)
Full inventory and install steps: [`docs/SKILLS.md`](docs/SKILLS.md).
| Workflow | Skill | Origin |
|---|---|---|
| Frontend | `frontend-design` | [anthropics/skills](https://github.com/anthropics/skills) |
| Frontend | `web-design` | [KAOPU-XiaoPu/web-design](https://github.com/KAOPU-XiaoPu/web-design) |
| Frontend | `vercel-react-best-practices`, `vercel-composition-patterns` | [vercel-labs/agent-skills](https://github.com/vercel-labs/agent-skills) |
| Frontend | `animate` | [emilkowalski/skill](https://github.com/emilkowalski/skill) |
| Frontend | `impeccable` | [pbakaus/impeccable](https://github.com/pbakaus/impeccable) |
| Frontend | `avoid-ai-writing` | [conorbronsdon/avoid-ai-writing](https://github.com/conorbronsdon/avoid-ai-writing) |
| Frontend | `anti-ai-slop` | local skill, not published here |
| Backend | `secure-coding` | [securityreviewai/secure-coding-skill](https://github.com/securityreviewai/secure-coding-skill) |
| Backend | `bola-detector`, `auth-rbac-scaffold`, `injection-checker`, `openapi-hardener`, `api-security-review`, `security-test-generator` | [apisec-inc/apisec-skills](https://github.com/apisec-inc/apisec-skills) |
| Architecture | `system-design`, `c4-architecture`, `isaqb-architecture-governance`, `secure-architecture-governance` | [Kotivskyi/architecture-governance-skills](https://github.com/Kotivskyi/architecture-governance-skills) |
| README | `readme-crafter` | [linhai0872/readme-crafter-skill](https://github.com/linhai0872/readme-crafter-skill) |
| README | `good-readme` | [adewale/good-readme](https://github.com/adewale/good-readme) |
| README | `curating-readme` | [liang-senbei/curating-readme](https://github.com/liang-senbei/curating-readme) |
## The QA harness
The [`qa/`](qa) directory keeps the workflows honest:
```bash
cd qa
python3 wf_quality_harness.py # score each workflow on 30+ criteria
python3 test_quality.py # gates: no regression, referenced skills exist, every phase verified
python3 test_tests.py # meta-tests: corrupt harness/workflows, confirm the score drops
```
The harness scores `wf-frontend`, `wf-backend`, `wf-architecture` and
`wf-security-review` (not `wf-readme`) by static criteria extracted from each
`SKILL.md`: structure, numbered steps per phase, verification verbs, referenced
skills that exist, and AI-isms density. The meta-tests are the useful part:
they deliberately break the harness and a workflow, then confirm the score
changes. If it did not, the detector would be useless.
Measured scores
Final scores per workflow come from [`qa/snapshots/baseline-final.json`](qa/snapshots/baseline-final.json).
The "before" values come from earlier harness runs; only the final snapshot is
versioned.
| Workflow | Before | Final | Gain |
|---|---|---|---|
| `wf-frontend` | 201.0 | 385.5 | +92% |
| `wf-backend` | 215.0 | 424.0 | +97% |
| `wf-architecture` | 199.9 | 413.7 | +107% |
| `wf-security-review` | 332.0 | 390.0 | +17% |
| Sum of the first three | 615.9 | 1223.2 | +98.6% |
| Sum of the four | 947.9 | 1613.2 | +70.2% |
The ruler also got stricter over time (13 criteria at first, 30+ later), so the
before/after comparison uses the final ruler on both sides. The numeric "2x"
target was not forced: inflating the score to hit it would be gaming the metric.
## Known limits
- The harness reads installed copies from `~/.hermes/skills/software-development//SKILL.md`
(hardcoded `SKILLS_DIR`), not the files in `workflows/`. Install in the directory
layout first, or it reports `SKILL.md não encontrado`.
- Two different skills share the name `api-security-review`: the vendored OWASP one
(`skills/`) and the apisec-inc one used by the backend workflow. Installing both
into the same directory makes one overwrite the other.
- There is no root CI workflow. `skills/hermaguard/.github/workflows/ci.yml` belongs
to the vendored hermaguard and does not run from this repository.
- The `wf-readme` workflow is not covered by the harness.
## Contributing
Open an issue first, then a PR. The rules are the ones the workflows enforce:
claims trace back to repo facts, every phase has a verification step, and docs avoid
AI-isms. Run the three `qa/` commands above before a PR; if you change scoring, update
`qa/snapshots/baseline-final.json`. Details in [`CONTRIBUTING.md`](CONTRIBUTING.md).
## License
[MIT](LICENSE), (c) 2026 Evandro Fonseca Junior. Vendored skills keep their own
licenses ([`skills/NOTICE.md`](skills/NOTICE.md)).