Skills para Shipping de Software
0xdeafcafe/skills
Automação completa de documentação e publicação de código
Ver no GitHubColeção de 21 skills Claude Code que automatizam tarefas repetitivas mas críticas do ciclo de desenvolvimento: ADRs, especificações Gherkin, audits, composição de PRs e iteração de reviews. Os skills se compõem em pipelines configuráveis, do planejamento até o merge.
Instalação
npx convez-ai 0xdeafcafe/skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar ADRs (Architecture Decision Records) retroativos ou preventivos
- Escrever especificações Gherkin e testes de caracterização
- Compor descrições de PRs e iterar sobre comentários de review
- Pipelines end-to-end de planejamento, implementação e publicação
Veredito

Stack sólida e opinada para quem quer delegar o trabalho chato mas necessário do shipping. Bem estruturado em skills compostas, com fluxos visuais claros. Excelente para times que querem documentação de verdade sem o atrito manual.
README
A tradução para o português deste README ainda está na fila.
skills
Twenty-one Claude Code skills for the parts of shipping software I'd rather not do myself. The ADR. The Gherkin spec. The audit pass. The PR description. The back-and-forth on review comments. Tedious enough to skip, important enough that skipping bites a quarter later.
They compose. Pick a front door.
Install
/plugin marketplace add 0xdeafcafe/skills
/plugin install skills@0xdeafcafe-skills
Shows up as skills @ 0xdeafcafe-skills in the plugin manager.
For other agents (Cursor, Codex, OpenCode) or selective installs, the skills.sh CLI:
npx skills add 0xdeafcafe/skills
The flow
flowchart TD
Start(["I want to work on something"])
Start -->|Known intent| StartFeature["/start-feature"]
Start -->|Exploring| StartDiscussion["/start-discussion"]
StartDiscussion --> StartFeature
StartFeature --> Branch{New or existing?}
Branch -->|New| Plan["/plan-change"]
Branch -->|Existing, ADR exists| Plan
Branch -->|Existing, no ADR| Backfill["/backfill-feature"]
Backfill --> Plan
Branch -->|Trivial| Skip([Skip scaffolding])
Plan --> Impl["/implement-change"]
Skip --> DriveChange["/drive-change"]
Impl --> DriveChange
DriveChange -->|Findings to fix| DriveChange
DriveChange -->|Clean| OpenPR["/open-pr"]
OpenPR --> DrivePR["/drive-pr"]
DrivePR --> Merged([Merged])
classDef entry fill:#fef3c7,stroke:#d97706,color:#92400e
classDef plan fill:#dbeafe,stroke:#2563eb,color:#1e3a8a
classDef drive fill:#dcfce7,stroke:#16a34a,color:#14532d
classDef ship fill:#fce7f3,stroke:#db2777,color:#831843
class StartFeature,StartDiscussion entry
class Plan,Backfill,Impl plan
class DriveChange drive
class OpenPR,DrivePR ship
/drive-plan is the composed alternative — it runs /plan-change →
/implement-change → /review-change end-to-end with phase-boundary
confirmations, for substantial changes that warrant the full pipeline
upfront.
tone-of-voice sits across the lot. It applies whenever Claude writes
under my name (commit messages, PR bodies, anything published) and to
its own conversational replies in this CLI, so the assistant stays in
my register rather than drifting into LLM tells.
What it looks like
A walked-through session, lightly compressed. Real artefacts get written along the way — ADRs, specs, characterisation tests, the PR body. The skills coordinate.
you: I want to add CSV export on the orders page.
claude: [/start-feature]
New code, or modifying something existing?
you: Existing orders page.
claude: No ADR for the orders module. Backing it up before we touch
it.
[/backfill-feature]
Reading src/orders/. Here's what I think the area does today:
- server-paginated listing
- cancellation through a state machine in OrderService
- no exports yet
Correct?
you: Yes.
claude: Writing retroactive ADR, characterisation spec, and 8 tests
locking current behaviour. All green.
[/plan-change]
Now the CSV export. What does the user click, and what comes
out the other side?
[discussion. ADR + spec drafted, reviewed, committed.]
you: Plan looks right. Let's build.
claude: [/implement-change]
Reading the spec. Plan: one new component, two edits to
OrderService, one new helper. 8 scenarios from the spec —
going to run them as smoke tests when I'm done.
[implementation back-and-forth, scenarios pass]
you: Done. Audit it.
claude: [/drive-change]
Triage: small (12 files, 2 directories). Skipping the slicer.
Fan-out to review-hygiene, review-code, review-test, review-feature,
review-security, review-ux in parallel.
review-hygiene: 230 prettier auto-fixable across 8 files (1 aggregate)
review-code: clean
review-test: 1 missing edge case (empty order list)
review-feature: clean
review-security: clean
review-ux: loading state missing on the Export button
Merger: 2 P1 findings, both mechanical. Dispatching to Sonnet
fix-appliers in parallel.
applied: 2 of 2
Re-run: clean. Ready for /open-pr.
you: Open it.
claude: [/open-pr]
...
PR opened: https://github.com/...
Drive it now, or wait for review?
The skills
| Skill | One-liner |
|---|---|
| Entry | |
start-feature | Front door for known work. Routes to plan-change or backfill-feature. |
start-discussion | Open-ended exploration. No file writes; talks until intent emerges. |
| Planning | |
plan-change | Drives the discussion that produces an ADR plus a Gherkin spec for what's about to be built. |
backfill-feature | Retroactive ADR, spec, and characterisation tests for existing code with no documentation. |
write-adr | Standalone ADR. MADR / Nygard / Y-statement, matched to whatever the repo already does. |
write-spec | Standalone Gherkin .feature file. |
| Reviewing | |
review-change | Read-only audit of the working tree. Slice → fan-out to specialists → merge → verify. Emits findings; never edits. |
review-pr | Read-only audit of an open PR. Terminal report by default; --comment posts findings as PR comments. |
review-hygiene | Runs project linters / formatters / LSP diagnostics in check-only mode. One aggregate finding per tool with auto: fix — fix-applier runs the tool directly, not the LLM. |
review-code | Per-file design judgment. SRP, modularity, length, naming, structural smells. Stays out of the linter's lane. Emits findings. |
review-feature | Feature logic against ADR / spec. Edge cases, error handling, side effects. Emits findings. |
review-test | Test quality on touched files. Right level, real assertions, no mocking the unit under test. Emits findings. |
review-security | Authz, secrets, input validation, dependency vulns, OWASP-top-10 smells. Emits findings. |
review-ux | Walks the changed UX in a real browser. Screenshots become finding evidence; a11y, console errors, network failures. |
review-spec | Audits a new spec or ADR against the existing corpus for duplicates, conflicts, overlap. |
| Driving | |
drive-plan | End-to-end orchestrator for substantial changes. Composes /plan-change → /implement-change → /review-change. Phase-boundary confirmations. |
drive-change | Workhorse for quick or clear changes. Implements from conversation intent + runs the agent pipeline (slice → fan-out → merge → verify) + dispatches fix-applier agents under sensitivity gating. |
implement-change | Translates a planned change (ADR + spec) into code. Runs the spec's scenarios as smoke tests. Hands off to /drive-change. |
| Shipping | |
open-pr | Composes title and body, runs final checks, opens via gh pr create. Wires tone-of-voice so the body sounds like the author. |
drive-pr | Iterates the open PR until trusted comments are resolved and CI is green. Calls /review-pr upfront, dispatches comment-driven edits through the agent pipeline. |
| Cross-cutting | |
tone-of-voice | Ghost-writes in my voice. Applies to PR bodies, commit messages, and Claude's own conversational replies. |
The agent pipeline
The aggregators (/drive-change, /drive-pr, /review-change,
/review-pr) compose four agent prompts in agents/:
orchestrate-slicepartitions a multi-file diff into domain-coherent slices and names the symbols that cross slice boundaries (LSP-backed when available, structural fallback otherwise).orchestrate-mergetakes parallel reviewer outputs, validates every finding against the JSON schema, runsgit apply --checkon the proposed fix, deduplicates byfile:line, and partitions into single-file work packets with sensitivity annotations.orchestrate-verifycross-checks the contracts the slicer named, catching the drift where slice A's change to interface I doesn't match how slice B uses I.fix-appliertakes one work packet, applies the fixes, runs a post-apply parse check, and reverts on failure.
Shared contracts live in references/:
finding-format.md+finding-format.schema.json— the structured shape every reviewer emits; the merger validates against the schema and silently discards malformed findings.change-envelope.md— the input shape every reviewer receives, the same shape whether reviewing a whole diff or one slice.language-tooling.md— when to prefer LSP queries (tslsp,gopls,pyright,rust-analyzer) over reading files.sensitivity-paths.md— path patterns that route fix-applier dispatches to Opus instead of Sonnet (auth / crypto / IPC trust boundaries).
Sensitivity gating means a Sonnet worker applies hygiene fixes in parallel across the codebase, while anything touching session handling, token storage, or IPC handlers routes to Opus regardless of severity.
The trust gate
PR comments come from anyone with a GitHub account. If a skill follows their instructions, anyone with a GitHub account has shell on my laptop. Every skill that reads them runs the same filter:
- AI bots — three trusted by name (CodeRabbit, GitHub Copilot
reviewer, Kilo Code reviewer). Other
[bot]handles are untrusted by default. - Humans — verified members of the repo's owning organisation,
or collaborators with
writeor higher, checked live viagh api. "Looks legit" isn't a verification.
Untrusted comments get read for context. They never move code, write a reply, or resolve a thread.
Six skills that read comments each ship their own copy of
references/trust-policy.md so a standalone CLI install of one skill
still has the policy locally. The canonical lives in
skills/drive-pr/references/trust-policy.md; the others are mirrors.
After editing the canonical, propagate it manually:
for d in review-hygiene review-code review-feature review-test review-security review-ux; do
cp skills/drive-pr/references/trust-policy.md skills/$d/references/trust-policy.md
done
Then stage the five mirrors. No script lives in this repo on purpose — skills with zero executable surface are safer to install.
Layout
.
├── README.md
├── .claude-plugin/marketplace.json
├── skills.sh.json
├── agents/ # prompt files invoked by orchestrators via Task
│ ├── orchestrate-slice.md # partition diff into domain slices
│ ├── orchestrate-merge.md # schema-validate + apply-validate + dedup + sensitivity
│ ├── orchestrate-verify.md # cross-slice contract check
│ └── fix-applier.md # single-file work packet executor
├── references/ # shared cross-skill contracts
│ ├── finding-format.md
│ ├── finding-format.schema.json
│ ├── change-envelope.md
│ ├── language-tooling.md
│ └── sensitivity-paths.md
└── skills/
├── start-feature/
├── start-discussion/
├── plan-change/
├── backfill-feature/
├── write-adr/
├── write-spec/
├── review-change/
├── review-pr/
├── review-hygiene/
├── review-code/
├── review-feature/
├── review-test/
├── review-security/
├── review-ux/
├── review-spec/
├── drive-plan/
├── drive-change/
├── implement-change/
├── open-pr/
├── drive-pr/
└── tone-of-voice/
Each skill is self-contained. When the CLI pulls a single skill on
its own, the recipient gets that skill's SKILL.md plus its
references/ folder. Cross-skill contracts in the top-level
references/ and the agents/ directory only matter to the
orchestrator skills that consume them; the standalone read-only
specialists don't depend on them at install time.
Reference files in each skill's references/ are loaded on demand,
which keeps the main SKILL.md lean.