think

name: think description: Use when there's no ticket or PR yet and the user wants to think through a problem, weigh approaches, or explore an idea. Trigger on phrases like "let's think about", "I want to think through", "let's brainstorm", "let's explore", "should we", "before we build", "no ticket yet", "I have an idea", "weigh A vs B", "what's the right approach for", "research X". Anchors to a vault initiative, captures decisions/scope/learnings passively into vault DB during dialogue, ends on a decision menu (create tickets / work existing / write hybrid plan file / defer). Use INSTEAD OF superpowers:brainstorming in Akshat's dev workflow.

For all per-ticket state mutations, see shared progress policy.

Think — Anchored Brainstorm

Open-ended exploration that anchors to a vault initiative, captures into the brain as you go, and ends on a clear next step. Borrows discipline from superpowers:brainstorming; diverges on three mechanics:

• Bundled questions via AskUserQuestion (not one-at-a-time).
• Plan file output at ~/.claude/plans/<slug>.md (not docs/superpowers/specs/).
• Vault passive capture during dialogue (3-way classifier into initiatives/<slug>/{decisions,charter,learnings}.md).

Wrapper owns capture — never invoke superpowers:brainstorming, never write outside wiki/ and ~/.claude/plans/.

Pre-entry: resolve initiative (mandatory, ≤3 steps)

Determine the initiative slug. Try in order, stop at first hit. Surface one line: ↳ initiative: <slug> (resolved via <step>).

1. Branch's existing progress.md context — if branch has an ENG-\d+ ticket, run python3 ~/.claude/scripts/progress_fm.py get <TICKET_ID> --field initiative. If non-empty, use it.
2. Keyword match — grep ~/opensource/vault/wiki/projects/*/learnings.md for ## Initiative: headings, match user's request against the slug + first body line of each. Surface top match: ↳ matched <slug> — proceed with this? (y/n).
3. Propose new — derive slug from user's request topic (lowercase, dashes, max 4 words). Ask once: ↳ creating new initiative <slug> under <repo>? (y/n). Repo defaults to cwd resolution; ask if cwd is not a repo.

Then load context: read the ## Initiative: <slug> section of ~/opensource/vault/wiki/projects/<repo>/learnings.md (skip if missing).

Body — bundled-question dialogue

Dialogue pattern

1. Reflect framing — restate what the user is exploring in one sentence so they can correct early.
2. Surface what's already known — pull from loaded initiative context. Cite specifically: per decisions.md: "...", per charter constraints: "...", per learnings.md: "...".
3. Bundled clarifying questions via AskUserQuestion — issue 2-4 questions in a SINGLE turn. Cover: constraint, deadline/reversibility, who's affected, success criteria. Multiple choice preferred over free text. NEVER ask one-at-a-time.
4. Propose 2-3 approaches — each with the main tradeoff named. Don't recommend yet.
5. Narrow with another bundled AskUserQuestion if discrimination is needed (1-3 questions max).
6. Recommend — pick one with main tradeoff explicitly named.
7. Capture — passive, see classifier below.

Iterate steps 3-6 as the user pulls threads. Don't railroad to a one-turn conclusion.

Passive capture — 3-way classifier

The moment a discrete insight crystallizes, route by content type and surface the destination filename:

Format for decision entries (append under ## Initiative: <slug> in learnings.md, never overwrite):

### YYYY-MM-DD — <one-line decision>
- Why: <rationale, one sentence>
- Trade: <what we give up, optional>
- Source: this /think session

Format for scope/constraint entries — append a bullet under ### Scope / ### Out of scope inside the initiative section; create the H3 if missing.

Format for learning entries — append a dated bullet inside the initiative section:

- YYYY-MM-DD: <past-tense insight>. Why it matters now: <one-line>.

NEVER ask "should I save this?". Save and notify inline.

Self-review trigger (>2 decisions in this session)

After the brainstorm produces a 3rd decision in decisions.md for this session, BEFORE the decision menu, run a quick scan:

• Contradictions — do any decisions in this session contradict each other or earlier decision entries in the initiative section?
• Ambiguity — could any decision be interpreted two ways? Pick one and make explicit.
• Scope drift — did decisions creep beyond the charter? Flag for user.

Surface findings in 3-5 lines. Fix inline (with user confirmation) before menu. <3 decisions → skip self-review.

Post-exit: decision menu

When user signals convergence ("ok let's do this", "I think we've got it", "let's move forward"), present:

Brainstorm complete. <N> decisions captured · initiative: <slug>

Pick next:
1. Create tickets   → /create-jira-ticket-with-reference
2. Work on existing → /work-on-jira-task <ticket-id>
3. Write plan       → ~/.claude/plans/<slug>.md  (this skill writes it directly — hybrid template below)
4. Defer            → already saved; come back any time

Never auto-pick. Wait for user.

Plan file template (option 3 — hybrid spec + plan)

When user picks option 3, write a plan file at ~/.claude/plans/<slug>.md. Same directory as plan-mode plans. Hybrid: spec-level rationale (so future-Claude understands why) + plan-level steps (so it can execute). Pull this session's vault captures inline.

Do NOT delegate to /writing-plans — this skill has already done the question-asking work; /writing-plans would re-ask. /writing-plans is for plans authored outside a brainstorm.

Template:

# <Initiative Title> — Plan

> Generated from /think session on YYYY-MM-DD. Initiative: `<slug>`.
> Captures from `wiki/projects/<repo>/initiatives/<slug>/{decisions,charter,learnings}.md` (this session).

## Context

<2-3 sentences: why this work, what prompted it, intended outcome.>

## Decisions (this session)

- **<decision 1>** — <rationale, one line>. Trade: <what we give up>.
- **<decision 2>** — ...
(One bullet per entry written to decisions.md during this brainstorm. Cite verbatim.)

## Constraints (from charter)

- <constraint or scope item>
- Out of scope: <X>
(Pulled from charter.md; only items relevant to this plan.)

## Prior Learnings Honored

- <past-tense insight that shaped the approach>
(Pulled from learnings.md; only those that influenced a decision above.)

## Approach

<Chosen approach in prose, 1-3 paragraphs. Reference decisions and constraints by name.>

### Alternatives Considered

- **<alt A>** — rejected because <one-line>.
- **<alt B>** — rejected because <one-line>.

## Critical Files to Modify

| Path | Action |
|---|---|
| `<path>` | <create/edit/delete> — <one-line what changes> |

## Sequencing

1. <step — small, verifiable>
2. <step>
3. ...

## Verification

<How to test end-to-end. Specific commands, files to inspect, expected outputs.>

## Out of Scope

- <explicit non-goal>
- <deferred item with reason>

Behavior on option 3:

1. Draft the file contents using the template, populated from this session's vault captures.
2. Show the draft to the user inline (or summarize sections if very long).
3. On user approval: write to ~/.claude/plans/<slug>.md. Surface: ↳ plan written: ~/.claude/plans/<slug>.md.
4. On user changes requested: revise inline; re-show; write on approval.

The plan filename is <slug>.md. If multiple plans per initiative, use <slug>-<topic>.md.

When NOT to use

• User shares a ticket URL → /work-on-jira-task directly.
• User shares a PR URL → /explain-anything or /get-pr-ready-to-merge.
• User asks for a tech doc → /create-tech-doc.
• User has a concrete bug to fix → superpowers:systematic-debugging.

Notes

• Project context update happens throughout, not just at end — compaction-safe.
• The wrapper exists specifically to prevent superpowers:brainstorming from creating side artifacts (docs/superpowers/specs/). If tempted to invoke that skill from inside this one, don't.
• Bundled questions are deliberate divergence from superpowers:brainstorming. The user prefers fewer turns over its "one question at a time" rule.
• Self-review at >2 decisions is deliberate — single-decision brainstorms don't need it; many-decision brainstorms drift without it.

Data Contract

Reads (DB)

• ~/opensource/vault/wiki/projects/<repo>/overview.md — project context (light)
• ~/opensource/vault/wiki/projects/<repo>/decisions.md — existing project-level decisions
• ~/opensource/vault/wiki/projects/<repo>/learnings.md — initiative sections (## Initiative: <slug>) for prior decisions / scope / learnings

Reads (Memory)

• ~/opensource/vault/wiki/projects/<repo>/progress/<TICKET_ID>/progress.md — for initiative resolution from branch (via progress_fm.py get --field initiative)
• ~/opensource/vault/wiki/projects/<repo>/open-threads.md — parked questions in this area

Writes (Memory + DB)

• ~/opensource/vault/wiki/projects/<repo>/learnings.md — passive capture (decisions, scope/constraints, past-tense insights) appended under ## Initiative: <slug>
• ~/opensource/vault/wiki/projects/<repo>/open-threads.md — append H2 if exploration parks a question (per CLAUDE.md)
• ~/opensource/vault/wiki/projects/<repo>/log.md — per-vault-v1: per-project log; append on session end: <ts> think: <slug> — N decisions, M new threads, terminal=<menu choice>. Fall back to ~/opensource/vault/wiki/log.md (vault-root) if the /think session is cross-project meta work with no resolved repo.
• ~/.claude/plans/<slug>.md — only if menu option 3 chosen; written directly by this skill using the hybrid template above (NO /writing-plans handoff)

DB writes from this skill are user-anchored captures during dialogue (not derived) — as durable as a manual Obsidian edit. Distinct from /brain-ingest's weekly distillation.

Local (skill-only)

• (none — wrapper deliberately avoids creating thoughts/, brainstorms/, side files)

Live external (not stored)

• Jira / GitHub MCP for resolution