# Dev handoff — brainstorm-with-docs

Resume context for a fresh agent working on **this skill's development**.
Dev-phase only; remove before the skill stabilises. For the backlog and decisions
see [ROADMAP.md](ROADMAP.md); don't duplicate it here.

## What this is

A custom, manually-invoked skill that fuses Matt Pocock's `grill-with-docs`
(glossary/`CONTEXT.md` + lightweight ADRs + active grilling) with obra/superpowers
`brainstorming` (explore → question → compare options → recommend → design doc),
tailored to the user. It is for thinking/alignment/documentation, never
implementation. See `SKILL.md`, `README.md`, and `artifacts/*.md` for the result.

Source files on disk (for comparison / not to be copied wholesale):
- `~/.claude/skills/grill-with-docs/{SKILL.md,CONTEXT-FORMAT.md,ADR-FORMAT.md}`
- `~/.claude/plugins/cache/claude-plugins-official/superpowers/5.0.7/skills/brainstorming/SKILL.md`

## State

- v1 shipped + reviewed. Commits `50640d8`, `f24b342` (pushed to origin/master).
- Repo path: `~/dotfiles/agents/.agents/skills/brainstorm-with-docs/`, symlinked
  into `~/.claude/skills/` (stow). Codex has a *separate* skills dir; this skill
  is not yet shared there.
- Cross-session memory: `project_brainstorm_with_docs.md` (in the agent's memory
  dir, indexed from MEMORY.md). The in-repo `ROADMAP.md` is now the canonical
  roadmap; memory is a pointer.

## Non-obvious rationale (the "why", not re-derivable from the files)

- **Frontmatter is portability-driven.** Claude reads all fields; Codex reads
  ONLY `name` + `description` (verified from Codex's own skill-creator). So extra
  fields are kept only when "useful on Claude, inert on Codex." `disallowed-tools`
  was removed — it's ineffective even on Claude (clears after the next message).
  `allowed-tools` space-separated is valid (Claude docs); it is pre-approval, not
  a restriction. Don't reopen this (= rejected finding F6 in ROADMAP).
- **The gate is the real control, not the harness.** User runs agents in auto
  mode, where permission prompts are unreliable. So `<hard_gate>` is written to be
  self-enforcing ("don't rely on the harness; announce before touching live
  systems") and allows investigation (read + read-only cmds + web research) while
  forbidding state changes.
- **`disable-model-invocation: true` is intentional** even though the user said
  "mostly manual" — for an *interview* skill, surprise auto-activation is more
  annoying than helpful; predictable manual invocation fits better.
- **Agile by design.** v1 = "boring-correct working hybrid"; the novel
  divergence mechanics are deferred to v2 on purpose, with a `<!-- SEAM -->` at
  `SKILL.md` step 5 kept self-contained so v2 slots in without restructuring.

## Open threads (decide with the user)

All tracked in `ROADMAP.md`. Most live ones:
- **R1** — restore grill's "give a recommended answer with each question" (likely
  accidental drop in v1). User leaning yes; deciding whether to do it now or after
  the first live test.
- **R2 / F1** — interview relentlessness + the divergence one-liner are held back
  on purpose so the first live test shows the pure-convergence baseline.
- **F4** — decide how `brainstorm-with-docs` and `grill-with-docs` coexist.

## Immediate next step

The user runs the **first live test on real tasks** next. Capture observations
(does it under-grill? do options collapse into one idea? is multi-context missed?)
and feed them into the R1/R2/F1 decisions. Append findings to this file or
`ROADMAP.md` as you go.

## Suggested skills

- **brainstorm-with-docs itself** — dogfood it; refining it is exactly a
  design/documentation problem.
- **write-a-skill** — sanity-check SKILL.md structure/frontmatter on changes.
- **grill-with-docs** — reference to check no valuable mechanic was lost (this is
  how R1/R2 were found).
- **conventional-commits** — required for any commit here (commits are in Russian,
  `type(scope): …` + Зачем/Что/Проверка body).
- **handoff** — refresh this file when a session produces a new stable state.