/scope. Assumption-First Scoping

Before doing a complex or irreversible task, restate the ask + flag assumptions + ask for corrections. Saves rework.

Source attribution: adapted from https://github.com/gsd-build/get-shit-done (MIT). Pattern: --analyze mode. The operator confirmed auto-trigger on CONFIRM-gate tasks.

Tone & Voice

Hard-edged, brief, no hedging in the restate. Read like a co-pilot confirming nav, not a customer-service rep. If an assumption is wrong, lead with the wrong one first. No "I think" / "maybe" / "perhaps". state assumption + reasoning.

Source & Pairs with

• the Autonomy Modes table in CLAUDE.md. the full tool→autonomy-mode mapping; consult before deciding if /scope auto-fires
• the helicopter-check lesson (list 3 ripple effects before any infra change). sibling pattern: 3 ripple-bullets before an infra change. /scope handles WHAT, helicopter-check handles BLAST-RADIUS
• CLAUDE.md Autonomy Modes section. defines AUTO / INFORM / CONFIRM / BLOCKED tiers
• Sister skills: skills/critic/SKILL.md (post-output adversarial review), skills-shared/sycophancy-guard/SKILL.md (catches drift in subsequent compliance), skills/agent-dispatch/SKILL.md (subagent briefing follows the same restate-pattern)
• outputs/raw/agent/ past /scope outputs for calibration

Example output (CONFIRM-gate restate)

## /scope restate

**Ask (as I read it)**: ship the notification engine v2 with reviewer routing + chat notify

**Concrete assumptions** (correct any):
1. "engine v2" = the modified version with weekly settlement (not the v1 monthly)
2. "reviewer routing" = route all notifications through a reviewer's inbox before you see them, not parallel-send
3. "chat notify" = personal chat, not a group topic
4. Trigger threshold stays at the $500 minimum per event
5. Storage: data/events.sqlite (existing), no schema change needed

**Out of scope**:
- Vesting logic (separate skill)
- Multi-party splits (v3 territory)
- Public-facing reporting (never auto-public)

**Blast radius**: 1 new daemon scheduler entry + 1 chat message-template addition + writes to the existing sqlite. Reversible by disabling the scheduler entry.

**Anything wrong?** I'll wait.

When to use

AUTO-TRIGGER. mandatory, no user invocation needed:

Any task classified as CONFIRM mode per the CLAUDE.md Autonomy Modes. That includes:

• Writes to your knowledge graph
• Writes to your CRM / spreadsheet
• Sending emails or outbound messages on your behalf (drafts staged, not sent)
• Modifying CLAUDE.md
• Creating / deleting scheduled tasks
• Any action visible to external parties
• Financial transactions
• Deleting files or data permanently
• Changes to daemon / MCP / infrastructure

In these cases, ALWAYS run /scope BEFORE the usual CONFIRM-gate ask. Don't just say "I'll write X to Y. OK?". first restate, flag assumptions, list out-of-scope, THEN ask.

AUTO-TRIGGER. heuristic:

• Task has 5+ assumptions (default values, scope boundaries, output format choices)
• Task is multi-hour and irreversible (code refactor, schema migration, mass data update)

USER-invoked:

• /scope
• "walk me through what you understand first"
• "before you build, restate"
• "analyze this"

DO NOT trigger for:

• Simple one-shot operations (read this file, edit this line)
• Quick answers to factual questions
• Tasks where the user explicitly says "just do it" / "go" / "build it" (override)
• AUTO-mode tasks (reading, searching, analyzing, answering)
• Re-runs of a task you've already scoped this session

Procedure

Step 1. Restate the ask

1-3 sentences, in your own words, not theirs. Capture the goal + the key constraint. Be concrete.

Example:

"You want a quotation spreadsheet that matches an existing template layout exactly, using the existing cost breakdown (34% markup, USD 6,500 anchor), shipped as a chat attachment tonight."

Step 2. List concrete assumptions

Bullet format. Concrete values, not hand-wavy. Include the default values you're picking.

Example:

• Management fee = 34% (matches the prior anchor, editable in F6)
• Payment method = one method only
• Single package tier
• Quotation No = the next in sequence (confirmed)
• Client total ≈ USD 6,507 (preserving the prior anchor)
• Payment split = 50% upfront / 50% post-delivery (from the prior deck)

Step 3. List what you're NOT doing

Explicit out-of-scope. Prevents scope creep.

Example:

• NOT modifying the internal costing breakdown content
• NOT re-sourcing vendors
• NOT re-running vendor pricing
• NOT touching the shortlist

Step 3.5. Helicopter view (mandatory, blast radius + ripple)

Zoom out before zooming in. 2-4 bullets, concrete:

• What this touches downstream: which other files / skills / crons / pipelines consume or depend on what I'm changing. Name them.
• Second-order effects: what behaves differently AFTER this lands that isn't obvious from the diff (a cron that now reads a new field, a hook that fires more, a memory file that grows every session).
• Reversibility: exactly how this is undone if wrong (revert one file / disable one scheduler entry / restore from backup). If NOT cleanly reversible, say so loud.
• Failure mode: the most likely way this bites later, and what would catch it.

This is the the helicopter-check lesson (list 3 ripple effects before any infra change) pattern promoted from infra-only into every /scope. /scope answers WHAT, the helicopter view answers WHAT ELSE MOVES. A restate without it under-serves the approval: the operator can only meaningfully approve what they can see the blast radius of. If the change genuinely touches nothing downstream (rare), say "blast radius: contained to {file}, nothing consumes it" explicitly. do not omit the section.

Step 4. Ask for corrections

One line:

"Correct any of these before I proceed?"

Step 5. Wait for the user

• "looks good" / "go" / thumbs-up → proceed
• Any correction → update the brief, re-confirm only if the correction is substantive
• Silence after 30 seconds on a clearly scoped ask → proceed on a good-faith interpretation

Anti-patterns

• Restate too vaguely ("you want a presentation". not specific enough)
• Skip the "NOT doing" list → scope creeps
• Hide assumptions in paragraphs → the user skims, misses the one wrong assumption
• Use scope mode for simple tasks → annoying, slows down
• Act before waiting for correction on CONFIRM-gate tasks → defeats the purpose entirely
• Restate but skip the assumption list → turns /scope into a summary, not a scoping exercise
• Skip the helicopter view (Step 3.5) → approval is uninformed. the user can't approve a blast radius they can't see

Example. when scope would have saved rework

A quotation-rebuild case: Without /scope, the agent defaulted the management fee to 25% and started building. It had to reverse-engineer to 34% mid-build to match the prior USD 6,500 anchor. Small save because the operator hadn't yet reviewed. but if they had reviewed at 25%, a full rebuild would have been needed.

Better flow with /scope:

1. Restate: "rebuild the quotation spreadsheet in the existing template layout, preserve the prior USD 6,500 anchor"
2. Assumptions: "34% management fee to match the prior anchor, single-method payment, single package tier, next-in-sequence quotation #"
3. NOT doing: "re-sourcing vendors, touching the shortlist, changing the cost breakdown content"
4. "Correct before I build?"
5. Operator confirms → zero mid-build reverse-engineering

Relationship to other skills

• CLAUDE.md autonomy modes: /scope is the execution mechanism for CONFIRM-gate tasks. AUTO tasks skip /scope. INFORM tasks usually skip unless ambiguous (5+ assumptions).
• Plan tool (ExitPlanMode): for architectural decisions (multi-file, multi-step design), use Plan. For scoping a single task, use /scope. Plan is broader-altitude.
• TaskCreate: /scope is about one task's brief. TaskCreate is about tracking multiple tasks. Both can coexist.
• Decision presentation: /scope is NOT a decision-presentation skill. It's a scoping skill. After /scope confirms, if the task also requires choosing between options, then present a trade-off table.

Why this skill exists

A session exposed the pattern: the agent was defaulting assumptions on CONFIRM-gate tasks without flagging them. A quotation rebuild + a code-adaptation task both had hidden assumptions that would have been easier to correct upfront than mid-flight. /scope codifies the "restate before building" discipline so the auto-trigger happens before the agent ever touches a knowledge-graph / CRM / CLAUDE.md write.