tech
/tech
Uruchamia trzech analityków równolegle: API, model danych i wymagania niefunkcjonalne, i szkicuje propozycje ADR do przebudowy produktu.
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 11.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Uruchamia trzech analityków równolegle: API, model danych i wymagania niefunkcjonalne, i szkicuje propozycje ADR do przebudowy produktu.
Oryginalny opis (EN)
This skill should be used when the user runs /rspec:tech or asks for the API, data model, non-functional requirements, architecture or ADRs of the product being reverse-specced. Runs three tech-analyst subagents in parallel (api, data-model, nfr) and drafts proposed ADRs for the rebuild.
Jak działa
SKILL.md bez nagłówków drugiego poziomu · 12 linii, 773 słów
Podgląd SKILL.md
---
name: tech
description: >
This skill should be used when the user runs /rspec:tech or asks for the API,
data model, non-functional requirements, architecture or ADRs of the product being
reverse-specced. Runs three tech-analyst subagents in parallel (api, data-model, nfr)
and drafts proposed ADRs for the rebuild.
metadata:
version: "0.11.0"
---
# /rspec:tech
Works best at stage `stories`; allowed from `features-approved`. To add contracts for features triaged into scope later, use add mode (see `rspec:features` Triage), never a full rerun.
1. Read `<spec>/rspec.json`. The features in scope are the MVP rows, plus Later rows when `scope` is `full`. The stack ADR must fit `deploy`, the deployment target: name it in the ADR's Context, rule out stacks that do not run well there (e.g. on Cloudflare Workers prefer a framework with a first-class Workers adapter and a Workers-compatible data store such as D1, KV or Durable Objects; Node-only servers, native modules and local SQLite files do not fit), and say which targets the choice rules out. If `deploy` is empty, ask for it in the same question as the stack; under autopilot use "undecided" and park it. With `deploy` undecided, the stack ADR still recommends one deployment target that fits the goal (with the runner-up) and designs for it, so the architecture is as concrete as with a given target; handoff records the accepted target in `deploy`. Always draft a platform ADR for the chosen target: its limits (request time, memory, storage, connections, plan tiers) and the NFR targets sized to them. If `stack` is empty, ask the user for the intended stack in one question (they may answer "undecided"); under autopilot, use "undecided" and let the ADRs recommend one. When the user gave a stack, the stack ADR recommends it unless the spec rules it out, and says why if it does.
2. Unless `<contracts>/schemas/` already has files, first decide the entities of the features in scope from the features and research: one list of singular kebab-case names with their schema files (`issue` → `schemas/issue.schema.json`). Give the same list to the `api` and `data-model` agents so every `$ref` matches, with one field-name casing for both (camelCase unless `stack` says otherwise); an agent that needs another entity adds it and says so in its reply. Then launch `rspec:tech-analyst` subagents **in a single message** for `data-model` and `nfr`, plus `data-migration` in the modernize track; when `data-model` has returned, launch `api`, which builds on the written schemas (on resume, only the pending areas; skip to step 4 if only `adrs` is pending). Every agent reads `<spec>/domain.md`. Give each: the path of `<spec>/domain.md`, `goal`, `scope` (the feature IDs in scope), `deploy`, `track`, `strategy`, `code` and source repo paths (modernize or open source), product, `lang`, `focus`, `stack`, sources from rspec.json, the list of research files, `<spec>/features.md`, its area (the agent knows which files to write), and the absolute path of `${CLAUDE_PLUGIN_ROOT}`.
3. When they return, fill the `Operations` column of `domain.md`'s permissions table from the api agent's reply, then reconcile: every `$ref` in `openapi.json` points to an existing file in `<contracts>/schemas/`; every feature in scope appears in some `x-features`, unless its story has `ui_only: true`; property names use the one casing in both `openapi.json` and the schemas; NFRs cover each `focus` area. When `goal` is `production`: every operation in `<contracts>/openapi.json` and every input screen appears as an entry point in `<spec>/threat-model.md` (add missing threats yourself), every threat has a mitigation, and every NFR has a Verification. Modernize track: every old table in the mapping appears in `<spec>/data-migration.md` (migrated or "not migrated" with a reason), and every entity in `<contracts>/schemas/` that holds migrated data is a mapping target. Run the check script, which validates the JSON, every `$ref`, the `x-features` IDs, `x-roles` and enums against `domain.md`, and body fields against their `x-entity` schema.
4. Draft ADRs in `<decisions>/` from `${CLAUDE_PLUGIN_ROOT}/templates/adr.md`, status `Proposed`, numbered from the next free `NNNN`. One ADR per decision the original product made that the rebuild must also make, typically: data store, realtime/sync approach, auth, ID scheme, API style, search. Context = what the original does (tagged); Decision = recommendation for the user's stack and goal. Then read `<spec>/nfr.md` against each ADR: an NFR a recommended decision cannot meet (e.g. offline editing when the sync ADR has no offline support) is either changed to fit, with the release it applies from, or named in the ADR's Consequences as a conflict for the user. Leave no silent contradiction; under autopilot, park each conflict and each ADR a document says is still needed.
5. Run `python3 ${CLAUDE_PLUGIN_ROOT}/scripts/check_spec.py .`; fix errors.
6. Set `stage` to `tech`, run `python3 ${CLAUDE_PLUGIN_ROOT}/scripts/check_spec.py . --mark-tech` (it records which criteria the contracts were written from; a later change to the stories then shows `tech is stale`), add a line to `<changelog>`. Reply with the counts of operations, schemas and NFRs, taken from the files (never from the subagents' replies), the ADR titles, the 2–3 decisions the user should confirm, and next step `/rspec:audit` (then `/rspec:handoff`).
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · tammai/rspec | 86 | MIT |
| rawkode-academy/rawkode-academy | 100 | MIT |
| karloscodes/fusionaly-oss | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 89011eead3d5 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100bmad-architecture
Tworzy architecture.md z ADR i systematycznym pokryciem NFR, mapując każde FR i NFR z PRD na konkretną decyzję projektową, żeby agenci deweloperscy trzymali się wspólnego podejścia. - 90discover
Dzieli produkt na moduły, uruchamia eksplorację każdego z nich równolegle i scala wyniki w pliku z ekranami specyfikacji. - 100architecture-design
Projektuje architekturę nowego systemu z wymagań: liczbowe cele niefunkcjonalne, porównanie opcji, ADR, diagramy C4 i plan etapów. - 95create-tech-spec
Pisze specyfikację techniczną z kontekstem systemu, kontraktami API, modelami danych, obsługą błędów i planem testów. - 86audit
Sprawdza, czy odwrotna specyfikacja jest prawdziwa, kompletna i spójna, uruchamiając trzech audytorów i poprawiając najważniejsze uwagi. - 100adr-decision-drivers
Pomaga wskazać, udokumentować i zważyć siły wpływające na decyzje architektoniczne: atrybuty jakościowe, wymagania niefunkcjonalne i kompromisy, które trafiają do ADR.