Act Idea

---
name: act-idea
description: Use when a feature, idea or change request is proposed: "can we add X", "it would be good if", "change request". Checks what exists, offers options with a recommendation, and after the decision files an effort estimate as backlog item and task. Not for planning a decided block - use act-prepare.
---

# Take in an idea or change request

Runs in the main session — decision and priority are the human's. Applies `R-work-idea-first`
(concept before code) and `R-human-ask` (bundled questions, no silent decisions). **Builds
nothing**: ends with a backlog item and, if picked up now, a task; building happens afterward via
`act-prepare`.

1. **Keep the wording verbatim** — the later yardstick, not the assistant's summary of it.
2. **Check what exists first** (`explorer` role): does this partly exist already, which files
   would be affected, and — above all — does it conflict with a decision in
   `docs/project/decisions.md`? A conflict makes this an ADR revision, not a feature.
3. **Write a concept** at `docs/project/concepts/<topic>.md`: starting point with findings,
   **options** (description, pros, cons, effort each), **recommendation**. "Do nothing" is a real
   option, with its consequences.
4. **Put it to the human** as a question in the inbox (`python .act/scripts/entries.py new
   question <title>` — writes `kind: question` and `for: all` into the header), options
   labeled, recommendation marked. Nothing built, no task filed, before it's answered. Any other
   open decision that comes up while filing goes into the inbox too, or stays on the backlog item
   as `decision: open` under `inbox-decisions: at-start` (`R-human-ask`).
5. **Estimate for the chosen path only**: scope, missing tooling (library, MCP server, rule set,
   a skill that doesn't exist yet — its own step, not "along the way"), a rough task breakdown
   with goal and acceptance check, and an ADR in `docs/project/decisions.md`.
6. **Priority and timing**, both from the human — plus the orchestrator's **own** estimate next
   to it, from a different angle (pressure, dependencies, risk already in the codebase), not the
   same number out of politeness. More than one step apart: say where the gap comes from.
7. **File it** (`python .act/scripts/entries.py new <kind> <title>` for each): a backlog item
   under `docs/ai/work/backlog/` (topic, priority, timing, effort, link to concept and ADR); a
   task under `docs/ai/work/tasks/` only for what starts **now**; a journal entry under
   `docs/ai/work/ledger/`.

**Shortcut for small things:** a typo, a field name, one config line needs no concept or ADR —
file it straight, but say the shortcut out loud so the human can object. When in doubt, don't
shortcut: a needless concept costs half an hour, a missed conflict costs a rebuild.

## Limits

- No implementation here, not even "just a quick start" — that preempts the decision.
- No recommendation without a real alternative, and no quietly aligned priorities — the gap
  between the two numbers is itself information.
- No effort estimate without a stated assumption behind it.