git:20260907.ea3a6dc to git:20260907.7bdf411

25 added, 12 removed. Audit A to A.

---

name: planning

description: "High-level development planning — features, TDD, architecture review. Triggers on: plan, feature, architecture, design, TDD, strategy."

argument-hint: "<what to plan: feature, tdd, architecture>"

---

# Planning

## Prompt refinement before scoping

Use the current turn's Attune refinement hook result, or call

`prompt_refinement(action="status")` if none was supplied. Follow its shared

policy to gather material missing context and incorporate answers into the

working prompt. If the user asks only to polish a prompt, return that prompt

without entering plan execution. An opt-out suppresses optional refinement,

not clarification needed for the actual planning task. Never restart intake

for a user's answer or correction, and never duplicate questions already asked

by refinement. Enable/disable only on the user's explicit request; a one-time

skip does not change the saved preference.

This skill fallback reaches planning invocations. Automatic delivery across

all messages requires a host that runs the plugin hook or follows MCP server

instructions; installing a skill alone does not establish that behavior.

**IMPORTANT: Start your response with a context preamble.**

Call `help_lookup(topic="spec-engine", mode="preamble")` and

display the returned `preamble` text as a blockquote. Then

tell the user they can say "tell me more" for a step-by-step

guide, or answer the scoping questions below to proceed.

If the MCP call fails, fall back to:

> **Planning** — Helps you plan features, architecture, and TDD strategy before writing code.

High-level development planning and architecture design.

## Routes

| Subcommand | Action |

| ---------- | ------ |

| `feature` | Plan a new feature |

| `tdd` | Plan TDD approach |

| `architecture` | Architecture review |

## MCP Tools

| Tool | What It Does |

| ---- | ------------ |

| `research_synthesis` | Synthesize insights from source documents at a path to inform planning |

Use `research_synthesis` when the user needs to gather

context from a directory of files or docs before planning.

Pass the directory (or file) as `path`; optionally set

`depth` to `quick`, `standard`, or `deep`:

```

research_synthesis(path="<dir or file>", depth="standard")

```

## Scoping

Before asking, reuse the subject, goal, scope, and constraints already

supplied. For ordinary requests such as "Help me plan a website for a

flower shop", ask only the material unknowns; do not make the user classify

the request as a feature, TDD, or architecture exercise first.

**Codex:** use the available built-in question tool for suitable independent

unknowns, without waiting for the user to request a form. In Default mode,

use `request_user_input_async`, respecting its current schema and limits.

Read the `elicit` skill's **Host defaults** for answer retention,

corrections, cancellation, and the asynchronous lifecycle. When the host

delivers replies during the active turn, keep needed questions pending rather

than finishing immediately; otherwise resume when its reply arrives.

The Attune server route and widgets are experimental

in Codex, not prerequisites or automatic fallbacks. Honor a request for

conversation and avoid unnecessary questions when a useful plan can proceed.

+ **Claude:** call the built-in `AskUserQuestion` directly for suitable

+ independent unknowns, without waiting for the user to request a form, and

+ follow its current schema (1–4 questions, 2–4 options each, free text via

+ the built-in "Other"). Read the `elicit` skill's **Host defaults** for the

+ synchronous reply lifecycle, answer retention, corrections, cancellation,

+ and the typed fallback for controls the schema cannot represent. Ask a

+ subject or problem statement as a plain conversational question when it

+ has no honest predefined choices. The Attune widget and server route are

+ experimental on Claude; use them only when the user asks for a form.

+ Desktop rendering and keyboard behavior remain pending observation.

**Antigravity:** use the `elicit` host default: enhanced conversational

- prompts, with native forms experimental. Claude's existing mapping below

- is unchanged pending its own host-specific review.

+ prompts, with native forms experimental.

For engineering planning tasks, clarify only missing dimensions:

1. **Type**: "What kind of planning? Feature spec, TDD

approach, or architecture review?" Skip when the request implies the type.

2. **Subject**: Depending on type:

- Feature: "What feature? What problem does it solve?"

- TDD: "What behavior should the tests verify?"

- Architecture: "What system? Any specific concerns?"

3. **Scope**: "How deep? Quick outline or detailed plan?"

- **Compatibility surface (Claude and hosts not yet reviewed).** The **Subject** phrasing branches on **Type**, so don't

- batch all three — ask **Type** first (a single `AskUserQuestion`) when

- it isn't already given by the `<what to plan>` argument. Once the type

- is known, **Subject** (a textarea) and **Scope** (quick / detailed) are

- independent and open: gather *those two* as one form via the `elicit`

- skill, **preferring the rich widget surface** (`elicitation_render_widget`

- → `show_widget`) with the AskUserQuestion mapping as fallback. If only

- one dimension is open, ask it as a single question — never force a

- one-field form (the §4 batching rule).

+ **Order on every host.** The **Subject** phrasing branches on **Type**,

+ so never batch all three. On Claude, ask **Type** and **Scope** (both

+ select-shaped and independent) in one `AskUserQuestion` call when both are

+ open, then ask **Subject** with the type-specific phrasing —

+ conversationally unless the conversation supplies honest candidate

+ options. On hosts not yet reviewed (Gemini is deferred), settle **Type**

+ first, then gather **Subject** and **Scope** as one form via the `elicit`

+ skill's compatibility path. If only one dimension is open, ask it as a

+ single question — never force a one-field form (the §4 batching rule).

## Execution

- 1. Use `EnterPlanMode` to create a structured plan

+ 1. Use `EnterPlanMode` to create a structured plan; on Claude, settle

+ remaining material choices before leaving plan mode — `AskUserQuestion`

+ for select-shaped ones, a typed question otherwise, and never against a

+ stated conversation preference

2. If context from multiple files is needed, call

`research_synthesis` first to gather insights

3. Present the plan for user approval before any

implementation