relay-plan
/relay-plan
Planuje projekt lub partię dla workera: czyta SPEC.md, przygotowuje szkielet i testy, a potem tworzy małe zadania i plan statusu na nowej gałęzi.
- Autor
- theasad13 ↗
- Repozytorium
- claude-code-token-saver ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 05.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 05.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Planuje projekt lub partię dla workera: czyta SPEC.md, przygotowuje szkielet i testy, a potem tworzy małe zadania i plan statusu na nowej gałęzi.
Argumenty
/relay-plan [path to spec file OR your idea]
Oryginalny opis (EN)
Relay Mode architect step. Reads an existing SPEC.md (or turns the user's idea into SPEC.md), builds the scaffold, interfaces and test setup, then writes a batch of small worker tasks plus relay/STATUS.md on a new git branch relay/<batch-name>. Use when the user says relay-plan or asks to plan a project or batch for the worker in Relay Mode.
Jak działa
- Steps
Podgląd SKILL.md
---
name: relay-plan
description: Relay Mode architect step. Reads an existing SPEC.md (or turns the user's idea into SPEC.md), builds the scaffold, interfaces and test setup, then writes a batch of small worker tasks plus relay/STATUS.md on a new git branch relay/<batch-name>. Use when the user says relay-plan or asks to plan a project or batch for the worker in Relay Mode.
argument-hint: "[path to spec file OR your idea]"
model: opus
effort: high
---
# relay-plan
You are the **architect**. A worker does the tasks later: Qwen Code / Kimi Code started by the USER by hand, or an automatic worker started by `relay-run` (Sonnet `relay-worker` subagent, or Codex CLI).
**Never launch an external worker CLI or call a model from a script** (no `qwen -p`, no `kimi`, no API scripts). Only the `relay-run` skill may start a worker (the Sonnet `relay-worker` subagent, or Codex CLI through `~/.claude/relay-kit/codex-worker.ps1`).
The user may not be a developer. Reply in the language the user writes in, in plain words.
**Token-saving rule:** No commentary or filler between tool calls. Do not re-read files you already read. Do not paste file contents back to the user. Talk to the user only when you must ask a question, and at the final step (the user's language, plain words).
Input from the user: $ARGUMENTS
## Steps
1. **Kit check.** If `relay/WORKER.md` is missing, do the relay-init steps first (run `~/.claude/relay-kit/relay-init.ps1`).
2. **Spec.**
- **If `SPEC.md` (or a spec file named in the input) exists:** read it. Do not re-ask anything it already answers. Make a short list of gaps. For each gap, pick a sensible default and write it under "Decisions and answers". Ask the user only about gaps that would force a task to guess (max 3 short questions, with options). Make sure the spec is saved as `SPEC.md` in the repo root and includes the **full test command**. Then go to Step 4.
- A spec from ChatGPT/Gemini describes WHAT to build. **You own HOW**: stack, modules, interfaces, data model and tests. If its technical choices are weak, too complex, or missing, replace them with your own and record each change under "Decisions and answers" as "(architect: <reason>)". Never copy a weak architecture just because it is written down. Treat "Suggestions for the architect" as optional ideas. Answer every item under "Open questions" with a default or, if it would force a guess, one of your max 3 questions.
- **If no spec exists:** ask only what you really need (max 5, short, with options): what it does, who uses it, must-haves for this batch, platform (web / CLI / mobile), and hard limits. Pick sensible defaults for everything else. Don't ask technical questions the user can't answer.
3. **SPEC.md (only if it did not exist).** Fill `~/.claude/relay-kit/templates/SPEC.md` into the repo root. Record every answer and default under "Decisions and answers". Include the **full test command**.
4. **Branch.** Pick a short batch name (e.g. `mvp`, `login`).
- If the repo has no commits yet: commit the kit first (`git add relay QWEN.md AGENTS.md` → `chore: add relay kit`).
- Then `git switch -c relay/<batch-name>`.
5. **Scaffold (you do this, not the worker).**
- Pick the simplest stack that fits. Install ALL dependencies the batch needs now, because the worker may not install anything.
- Add a `.gitignore` (node_modules, .env, build output, etc.).
- Write the interfaces/contracts as stubs (function signatures, types, TODO bodies).
- Set up the test runner with one passing smoke test. Run it and confirm it works.
- Prefer that **you** write each task's tests now (the worker can't change them). The tests fail until the worker implements the code. That is expected.
6. **Tasks.** Split the rest into small tasks in `relay/tasks/TASK-001.md`, `TASK-002.md`, … using `relay/tasks/_TEMPLATE.md`:
- 1–3 target files each, with exact paths. Clear contracts and acceptance criteria. 2–3 read-only context files.
- `test_command` runs only that task's tests, with a verbose reporter (jest `--verbose`, vitest `--reporter=verbose`, pytest `-v`, node `--test` prints names). Point it at exact test files, not whole folders. `max_retries: 2`.
- `expect_output`: the exact name of one test of that task. verify.ps1 FAILS if this name is not in the output. This stops "0 tests ran = PASS". Never leave it empty.
- Do not wrap the whole command in quotes, and do not add `2>&1` or other redirection; verify.ps1 captures output itself.
- Each task must be finishable without guessing. If something is unknown, ask the user now.
7. **STATUS.md.** Fill the batch name, branch, full test command, and one row per task in dependency order (`PENDING`, attempts `0`).
8. **Sanity check.** Run `powershell -NoProfile -ExecutionPolicy Bypass -File relay/verify.ps1 -TaskId TASK-001`. It should run the tests and parse the header. FAIL because of missing code is fine; a broken header or command is not. Also confirm the task's tests really ran: "No test ran" or "Expected test ... did not appear" means the filter is wrong. Fix the header before committing.
9. **Commit** everything: `relay: plan batch <batch-name>`. Do not push.
10. **Tell the user** (their language, short): what the batch builds, the task list, any defaults you chose, and the three ways to continue. Always include the `/clear` reminder. Translate these lines into the user's language, and keep the commands exact:
> 💡 **Save tokens:** type `/clear` first. Everything is saved in files, so the old conversation is not needed.
> **Sonnet mode (automatic):** after `/clear`, type `relay-run`. The Sonnet worker does all tasks, then Claude reviews them.
> **Codex mode (automatic):** after `/clear`, type `relay-run codex`. Codex does all tasks, then Claude checks and reviews them.
> **Qwen mode (manual):** open Qwen Code (or Kimi Code) in this folder and type: `Do all pending tasks in relay/STATUS.md`. When it finishes, type `review` here.
Wersje
W bazie nie ma innych skilli o nazwie relay-plan.
Migawka bazy z 13.11.2026 · wersja paczki ec17206eeeed (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
- 90relay-run
Uruchamia automatycznie oczekujące zadania w pliku statusu, jedno po drugim, przez agenta-workera, a potem przegląda partię z ograniczoną liczbą poprawek. - 90relay-review
Przegląda pracę workera po partii: raporty zadań, weryfikację, zakres zmienionych plików, testy, jakość i bezpieczeństwo, a potem wystawia uwagi do poprawy. - 100test-python-binding
Buduje i testuje wiązanie Python do NeMo Relay oraz wtyczkę workera: zmiany w python/nemo_relay, python/plugin i crates/python. - 90relay
Przetwarza partię poprawek użytkownika jako ścisłą pętlę pracy: segregacja, jedna pozycja naraz, sprawdzone kryteria; nowe uwagi nie przerywają pracy. - 100writing-plans
Tworzy plan z zatwierdzonego projektu lub specyfikacji: małe zadania test-first z plikami, interfejsami i oczekiwanym wynikiem. - 95relay-fragments-patterns
Opisuje wzorce Relay dla aplikacji React: kompozycja fragmentów, maskowanie danych, kolokacja i wzorce kontenerów.