feature-discuss

name	feature-discuss
description	Capture implementation decisions for a feature before planning begins — libraries, error handling, edge cases, scope boundary. First step of the full feature loop.
disable-model-invocation	true
allowed-tools	Bash, Read, Write

<runtime_note> Explicit-invocation only (/feature-discuss). Never auto-triggers. </runtime_note>

Capture how to build the feature, not just what to build, before any planning happens. Planning cannot start on solid ground until these decisions exist. Output is .context/features/NNN-slug/CONTEXT.md.

Shared conventions — directory layout, file ownership, task statuses, in-flight definition, feature numbering, branch and commit policy — live in references/context-contract.md in the feature skill's directory (installed at ~/.claude/skills/feature/references/context-contract.md); follow it rather than re-deriving.

1. Ground yourself

OVERVIEW.md and STATE.md are already in context — the project CLAUDE.md's @imports load both at session start. Re-read them directly only if they changed during this session, or if the imports are missing (no CLAUDE.md, or it lacks the @imports line). Don't read research/ or DECISIONS.md in full — unless the conversation surfaces a specific reason to check further. Keep this step cheap.

2. Scope check (borrowed from brainstorming's discipline)

Before asking anything else, ask yourself: does this request actually describe multiple independent features? (E.g., "add auth, billing, and an admin dashboard" is three features, not one.) If so, say this to the user directly and recommend running /feature-discuss separately for each, starting with the first. Don't spend the Q&A below refining a request that needs splitting first.

3. Lightweight Q&A — assumptions mode by default

Propose your best guess at the implementation decisions, and ask the user to confirm or correct, rather than open-ended interviewing — this is faster in the common case where your guesses are reasonable. Cover, as relevant:

• Libraries / dependencies to use or avoid
• Error-handling strategy for this feature
• Whether behavior is scoped narrowly or applies globally
• Edge cases worth deciding now rather than leaving to the planner's judgment
• Anything from STATE.md or OVERVIEW.md that constrains the approach

Ask one question (or one batch of proposed assumptions) at a time. If the user says "just use your judgment" on something, note that explicitly in the output rather than silently picking — the planner should be able to tell "decided by user" from "Claude's default judgment call" at a glance.

If the user prefers a fuller interview instead of assumptions-mode for this feature, switch — don't insist on the faster mode against their stated preference.

4. Write CONTEXT.md

If this skill was invoked directly (no /feature router before it) and feature/NNN-slug doesn't exist yet as a branch, create it now and record the pre-branch HEAD as the Base SHA — the same thing /feature's step 3 normally does ahead of this skill:

git rev-parse HEAD    # only if no feature branch exists yet — this is the Base SHA
git checkout -b feature/NNN-slug
mkdir -p .context/features/NNN-slug

# Feature: <name>

**Base:** <starting commit SHA>

## Goal
[One sentence — what this feature does]

## Scope
[What's in scope. What's explicitly out of scope, if it came up.]

## Implementation Decisions
- [Decision]: [what was decided, and who decided it — user or Claude's default judgment]
- ...

## Global Constraints
[Project-wide conventions every task in this feature must respect — established patterns, style rules, invariants. This is what `prompt-engineer` restates verbatim to executor subagents that never see this file.]

## Open Questions
[Anything explicitly deferred to the planner or executor's judgment]

5. Hand off

Tell the user the next command is /feature-plan. Suggest --thorough whenever the feature introduces a dependency not already in the project's manifest, or is a first-in-this-repo integration (a new external service, protocol, or pattern this codebase hasn't used before) — either signals real research is needed before planning.