davide-pi/claude-dev-kit

A personal Claude Code toolkit for a developer who
has to cover every role alone: analyst, backend, DBA, frontend, reviewer, release manager, ops.
Opinionated about C# / ASP.NET, EF Core, SQL Server, PostgreSQL, Redis, RabbitMQ, TypeScript,
Angular, React, with Azure DevOps as the ALM platform and GitHub alongside it. Opinionated on
technique, anonymous on identity: the patterns, libraries and commands are real, while
organization and project names, machine paths and addresses are placeholders.
Italian by default; English for the code and for what the AI reads back. Everything an asset
makes a person read is Italian — chat reports, commit messages, pull request titles and bodies,
review comments, work items, wiki pages. English is reserved for the code, for a project's technical
documentation tree, for plan files and specs (a later session reads those back, which makes them
AI-facing reference), and for the instructional prose of the assets in this repository. So an asset
is written in English and says "produce Italian" — swap that second half for your own language and
nothing else changes. Machine identifiers are never translated in either direction.
## Five levels
| Level | What it solves | Assets |
| Routing | which asset owns this task | the route table in CLAUDE.md, path-scoped rules/ |
| Knowledge | not re-deciding what was decided once | 33 skills, depth in reference files loaded on demand |
| Exploration | finding context without burning the main context window | 10 subagents |
| Execution | running a known sequence in one shot | 8 one-shot skills (/name), 3 saved workflows |
| Safety net | not forgetting, and not doing the irreversible | 14 hook files (nudges.js runs four of them), the completion gate, the validator |
Two rules govern all of them, and they live in CLAUDE.md so they apply without being triggered:
route through the table, and try the CLI before the MCP server, every time.
## Skills
One skeleton throughout — When (with what it explicitly excludes), Decide (a decision table or
tree), Do (copy-pasteable PowerShell), Traps, References — a Contract. line after the title
(what it produces · what it never does · the output language), the description a Use when trigger of
25 words at most, the file capped at 120 lines, with depth in references/ files loaded only on demand. No skill pins a version: where the answer depends on one,
the skill says which file to read or where to verify it.
Routing and process. These absorb the parts of the superpowers plugin worth keeping, without
the ceremony and rewritten for PowerShell.
| Skill | Purpose |
| skills/plan-work/SKILL.md | A plan file with verifiable tasks, sized to the change — and when not to write one at all. Also turns a work item into an attack plan before any code. |
| skills/deliver/SKILL.md | Every code change from triage to verdict: scenario and size, a doubts pass before the loop on M and L (a question only when its answer changes behaviour), a written perimeter (per worktree, in the git dir), acceptance criteria with their proof (scaled by size) before the edit, the impl-review-loop workflow once per coherent request (size S: the session edits inline, the loop only reviews), the mechanical scope-guard.js --check after an OK, tests by level run in the main session, a proof check of the criteria's tests, a verdict block (with the defaults chosen and a proposed local test) and a table of what was left outside. References for the scenarios and for perf baselines. |