brainstorm

/brainstorm — Idea → Design → Execution

A single skill for the full cycle: discussion → design → plan → autonomous execution.

Philosophy

• One approval point — after showing the design+plan. Then autonomously
• Questions only BEFORE the plan — do not ask permission for every step
• Skills-first — NEVER write your own scripts if a skill exists

Routing (automatic)

The AI determines the mode by the task itself:

Signs	Mode	What to do
1-2 objects, obvious implementation	express	Briefly state the plan → execute with skills
3-5 tasks, design needed	standard	Full brainstorm → plan → execute
6+ tasks, architectural	full	Brainstorm → plan → subagents with review

Process

Phase 1 — Understanding (dialogue with the user)

Load context:

rlm_route_context(query="project architecture current work")

For BSP-related tasks:

ssl_search(query="<ключевые слова>")

Ask clarifying questions one at a time via AskUserQuestion:

• Prefer answer options (not open-ended questions)
• Focus: why, who uses it, constraints
• 1C-specific: configuration or CFE? Which objects? Is BSP needed?
• Stop when the task is clear — do not ask for the sake of asking

Phase 2 — Design + plan (THE ONLY approval point)

Show the user in a single message:

## Дизайн: <название>

**Подход:** <1-3 предложения>

**Объекты:** <что создаём/меняем>

**Задачи:**
1. <задача> — скилл: /meta-compile
2. <задача> — скилл: /form-compile
3. <задача> — скилл: /subsystem-edit
...

**Риски:** <если есть>

Выполняю?

Scale to complexity:

• express: 3-5 lines, no risks section
• standard: full format above
• full: + alternative approaches with trade-offs, + a "Rejected options" section

Wait for a response. The user will say "ok" / adjust / decline.

Phase 3 — Autonomous execution

After "ok" — execute silently, without asking for permissions.

Express/Standard mode: execute yourself

For each task from the plan:

1. Call the corresponding skill (Skill tool)
2. If the task is BSL code, check via 1c-syntax-checker
3. Commit atomically (one task = one commit)

Full mode: subagents

For each task:

1. Dispatch implementer (Task tool) — give the full task text + list of skills
2. Dispatch reviewer — check the result
3. On ❌ from reviewer — fix and continue (do not ask the user)
4. On a blocker (impossible to continue) — then ask

Subagent prompts: .claude/skills/subagent-dev/implementer-prompt.md, .claude/skills/subagent-dev/reviewer-prompt.md

Phase 4 — Report

One final report:

Готово: <что сделано>
Файлы: <список изменённых>
Коммиты: <N коммитов>
Замечания: <если были проблемы в ревью>

Saving artifacts

For standard и full modes save:

• openspec/changes/<feature>/design.md — design
• openspec/changes/<feature>/tasks.md — plan with [x] marks

For express — do not create files, they are not needed.

PROHIBITIONS

• ❌ Asking “can I execute task N?” after the plan is approved
• ❌ Asking “does this part look correct?” section by section
• ❌ Writing PowerShell/BAT scripts if there is a skill
• ❌ Writing XML by hand if there is /meta-compile, /form-compile, etc.
• ❌ Generating BSL from memory without checking via 1c-help and 1c-syntax-checker