# Intent Checkpoint > Resolve consequential ambiguity about scope, outcome, owner, or next stage with a brief user decision. Use before unselected expansion, including optional protection work; not for clear reversible work or as a substitute for host approvals. - Skill: `seasoncake/intent-checkpoint` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add seasoncake/intent-checkpoint` - Raw SKILL.md: https://api.skillmd.com/api/skills/seasoncake/intent-checkpoint/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - License: MIT - Author: SeasonCake (https://skillmd.com/u/seasoncake) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/seasoncake/intent-checkpoint --- # Intent Checkpoint Turn a material fork into a small explicit choice, then continue the selected work. This is a decision workflow, not a new approval system. ## Decide whether to ask Compare the latest instruction with the proposed action: object, outcome, owner, stage and exclusions. Resolve factual uncertainty with a cheap in-scope check first; do not ask the user to diagnose for you. For forwarded advice or restrictions, establish who selected them and which owner/stage they govern. Delivery as a user message does not itself turn a sender's scope statement into a new author decision. - Clear selected endpoint: perform its normal reversible steps without repeated questions. - Material fork: ask when plausible interpretations change behavior, ownership, deliverables, substantial cost, or destructive/external/later-stage action. - Explicit actor/start reservation: retain it until superseded by a later instruction. A completed prerequisite is not the signal; do not dispatch a message to manufacture it. - Speculative adjacent idea: usually record or omit it, not a decision for every opportunity. Optional encryption, signing, anti-debugging or release gates are not selected by phrases such as "for safety", old experiments or audit suggestions. Distinguish a real failure of the existing baseline from a new protection proposal; neither hide the failure nor make the proposal mandatory. Copy simplification does not authorize removing underlying controls. ## Ask the smallest useful decision Use the user's language and level of detail. Prefer one question, or up to three closely related decisions. State verified context, alternatives and consequences; offer two or three genuine choices with a supported recommendation when useful. Include no-change/defer when real, respect free text, and avoid loaded safe/unsafe labels, false urgency, bundled permissions or a long questionnaire. Use an available native question tool only when current host instructions permit it for this purpose. Read [native forms](references/native-forms.md) when choosing/adapting the call. Otherwise ask one concise plain-text question. Do not invent tools, install plugins, build a form service or bypass host approvals to force a form. Pause only the affected action while awaiting the material choice; continue independent selected work. Do not resend the same form or poll an unchanged answer. If nothing else can proceed, explain the narrow unresolved choice and yield. ## Bind the response and continue Only a submitted answer or direct instruction resolves the choice. Preselection, accepted tool calls, timeout, closed cards or a host "resolved" event alone are not user consent. Correlate each answer with its original question and scope, not its arrival position. After compaction or delayed/replayed delivery, reconcile it with the latest substantive instruction and current stage. An old planning answer cannot undo a later explicit start or repeat completed work. Ask only if a genuinely new conflict remains. Apply local completion, blockers and follow-up questions only to their affected scope. Continue other selected work unless the user changes the goal or a verified shared dependency prevents it. Reaching the selected endpoint is a valid stopping point; describe the next stage without inventing its authorization. Record consequential decisions briefly in the existing plan/nearest record: selected action, object, stage and exclusions. No separate receipt system is required; historical records are retrieval cues, not perpetual authority. Continue through the selected endpoint without repeated permission checks, unselected proposals or inferred publication, deployment, deletion or another actor's start. Actual execution still obeys target checks, resource/recovery limits and host permissions. ## Evaluate the decisions When adapting/evaluating this skill, read [synthetic cases](references/synthetic-example.md). Check a material fork that should prompt and a clear instruction that should not, plus delayed-answer scope when relevant. Evaluate the action, not word matching. Structure, scenario review and actual automatic matching/host interaction are different evidence; forms do not guarantee prevention of incidents.