# Ank Plan > Interview a goal into decisions and tasks recorded in .ank/. Use when someone brings a feature, change, or problem to plan before implementation in a repository with a .ank/ directory. - Skill: `haksolot/ank-plan` (Agent Skill) - Install (CLI): `npx skillmds@latest add haksolot/ank-plan` - Raw SKILL.md: https://api.skillmd.com/api/skills/haksolot/ank-plan/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: haksolot (https://skillmd.com/u/haksolot) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/haksolot/ank-plan --- # ank-plan Planning turns an intention into entities the loop can consume: decisions as ADRs, rules as specs, work as tasks, order as blocked_by. The interview is how uncertainty leaves the plan before an executor meets it. The ank skill is the contract and applies here in full. This file adds the planning policy only. ## Read before asking Never ask a question the repository answers. Before the first question: ank context what already binds this perimeter ank find what already exists near the goal ank find --type spec the specification itself ank scope everything covering the files in play ank log what previous holders learned Then read the code the goal touches. A question answered by reading, asked anyway, teaches the human to stop answering. **Re-measure what you are about to plan around.** A number from an earlier session is a claim about a tree that has moved since, and a goal built on one is a goal aimed at the wrong place. Take the measurement again before the first question: the work may already be done, and where it is not, the shape of what remains is rarely the shape that was reported. ## The interview Work the frontier: the decisions whose prerequisites are already settled. Everything downstream of an open decision waits. For each decision on the frontier: 1. Check the repository has not already made it. 2. Recommend one option, and say why. 3. Ask. One round holds only questions independent of each other. Fold the answers in, recompute the frontier, repeat. Stop when the frontier is empty: nothing important left silently assumed. A question only running code can answer is not a question. It is a task with an experiment for a criterion. ## Write the plan into the corpus Each settled decision becomes exactly one kind of entity: a durable architectural decision ank new adr lands proposed a normative rule ank new spec lands proposed concrete bounded work ank new task a scope is mandatory an ordering between tasks ank amend --blocked-by a passing observation ank log A task names the policy its shape calls for with `ank new task --method `: `diagnose` when the criterion names a defect, `tdd` when it names behaviour to build. Leave it unset otherwise, since an absent method says only that nobody designated one. The name is the sibling's directory, never ank-tdd. A done_criteria states what a verifier can check, not what an executor should attempt. If the criterion will not write, the decision is not settled: back to the frontier. `new` is not where it is enforced — it writes a task with no criterion and says nothing — `claim` is, and it refuses at 7. So a criterion left for later is a task nobody can start, and the plan is short one decision. ## Where planning ends Proposed ADRs and specs bind nobody until a human ratifies them: a signed accept, on the default branch. Say what is waiting and stop. Then ank graph and ank check on the result: a cycle, an orphan or a finding is a plan not finished.