git:20260930.989655a to git:20261003.1365be2

13 added, 4 removed. Audit A to A.

---

name: skiphow-plan

description: Prepare an idea or substantial change for implementation with research, a reviewed specification, vertical slices, and an execution prompt. Explicit invocation requests durable planning records; ordinary read-only planning stays read-only. Stop before implementation. If context was compacted while this workflow governed, reopen this file and the SkipHow CTO skill before the next consequential action.

---

# SkipHow plan

Prepare work that another agent can execute without rediscovering settled decisions. Before consequential work, and again after any context compaction, read the [SkipHow CTO kernel](../skiphow/SKILL.md) unless its full text is already in context; a summary is not its text. It remains responsible for technical judgment and owner questions.

Explicit invocation of this skill requests a specification and planning records in the project's authorized tracker. Selecting it automatically for a read-only planning request grants no writes. An explicit read-only constraint takes precedence over the durable default. Keep implementation, commits, and delivery outside this workflow.

Use [product](../skiphow/references/product.md) to recover the intended outcome, current behavior, constraints, exclusions, and acceptance conditions. Investigate before taking genuinely unresolved product choices to the owner. Carry the interview through the questions their answers reveal until no known product choice remains open in the accepted execution scope. Preserve settled answers, keep dependent behavior undecided until its answer arrives, and prepare independent work meanwhile. Missing answers leave incomplete preparation, not an execution-ready plan; later discoveries can still require a decision during execution.

Use [technical design](../skiphow/references/technical-design.md) for consequential choices and build-versus-reuse analysis. Establish what the repository already provides before proposing new logic or dependencies. Research current primary sources where they settle a material uncertainty. Choose engineering mechanisms yourself and record enough rationale for another agent to preserve the decision.

For user-facing work, carry the relevant journey scenarios, design direction, and identified reuse constraints into the specification and tasks that depend on them. Separate accepted product decisions from proposals and engineering judgments. Preserve enough source context for an implementer to understand what must hold and why. Record the resulting decisions and acceptance conditions independently of the specialist capability that helped establish them; an executing agent reassesses the capabilities available on its host while preserving the accepted outcome.

- Apply [delegation](../skiphow/references/delegation.md) to substantial independent research and to slicing. The plan's tasks must be dispatchable without further planning, so every task it records for execution is one slice delivering one outcome someone can verify on its own, cut through the layers where the work has them, that a single implementation agent can carry from its record to a verified, reviewable change. A grouping record may collect tasks but is never itself dispatched. Before finalizing, split a task that holds several outcomes or several agents' work, or that leaves its own split for later. Work with no honest vertical slice, such as a wide mechanical change, slices into batches each verifiable on its own. Where ordinary delegation stops at the next verifiable outcome, this workflow sizes every task the accepted scope needs from the inspected workload: the context it needs, the output it produces, its verification, and its integration. Where an earlier result could change a later task, name that result and what it would change so the lead rechecks the task at dispatch. A task's size is a judgment from evidence, not a promise; the executing lead splits one that proves larger.

+ Apply [delegation](../skiphow/references/delegation.md) to substantial independent research and to slicing. Every task recorded for execution delivers one independently verifiable outcome, cut through the layers where the work has them, that one implementation agent can carry from its record to a verified, reviewable candidate. One assignment allows normal tool calls and review corrections. A grouping record collects tasks and is never dispatched. Size every task in the accepted scope now; split one that needs several agents or defers its own split. Keep implementation and its necessary checks together when separate parts have no independently verifiable outcome. Small cohesive work stays one task. Compatible migrations and mechanical work use independently verifiable units; a large uniform automated change may stay one task when the meaningful change is bounded and completeness can be checked. A task's size is a judgment from evidence, not a promise; the executing lead reslices one that proves larger.

- Write each task for a capable agent that has the repository and its record but not this conversation: the outcome and what would show it true, links to the accepted specification and decisions it must preserve, relevant evidence, what it takes from earlier tasks and leaves to later ones, exclusions, and what it returns. Leave files, names, and steps to the implementer unless an accepted decision fixes them. Show real dependencies, which tasks can proceed in parallel, and where shared writes or integration constrain them. Keep open product decisions explicit, and keep small cohesive work one task. Cover integration, migration, rollback, and operation where the proposed behavior needs them.

+ Where the accepted work lacks a proven path through its components, choose a thin first slice with real behavior. Establish only the structure and component connections that slice needs, verify the working path, then expand it through the remaining bounded slices. Reuse an established path in a mature project. The first working behavior preserves required safety, integrity, and compatibility. Name what its result unblocks and which later assumptions the lead must recheck at dispatch. Independent preparation can proceed before that result; prepare the whole accepted task set without inventing a universal foundation stage.

- Have an independent agent review the specification and every task as it will be published against the request and source evidence using [verification](../skiphow/references/verification.md). Check missing outcomes, conflicting decisions, and verifiable acceptance conditions; for each task, whether it holds one outcome that one agent could finish from its record alone; and across tasks, invented or missing dependencies, overlapping work, parts of the result no task covers, unnecessary implementation prescription while retaining justified contracts and reuse constraints, and whether the planned review and integration workload is itself bounded. Check whether the proposed journey fits its application context and whether consequential product choices have an authoritative basis. Resolve supported findings before finalizing the records. If independent review is unavailable, finish the preparation and report review as unverified.

+ Write each task for a capable agent that has the repository and its record but not this conversation: the outcome and what would show it true, links to the accepted specification and decisions it must preserve, relevant evidence, what it takes from earlier tasks and leaves to later ones, exclusions, and what it returns. Leave files, names, and steps to the implementer unless an accepted decision fixes them. Show actual blocking results, available parallel work, and shared writes, runtime resources, or integration capacity that constrain dispatch. Consuming the same settled API creates no dependency by itself. Name any earlier result that could change a later task and what to recheck at dispatch. Keep open product decisions explicit. Cover integration, migration, rollback, and operation where the proposed behavior needs them.

+ Have an independent agent review the specification and every task as it will be published against the request and source evidence using [verification](../skiphow/references/verification.md). Check missing outcomes, conflicting decisions, and verifiable acceptance. A shared outcome alone does not establish one-agent feasibility: assess required context, materially different changes, manual implementation, verification, review, and integration together. Where workload is materially uncertain, use existing measurements, comparable completed work, or representative source and check inspection; a full inventory is not required. A size finding names the concrete load and a smaller independently verifiable cut that preserves all acceptance and dependencies. Split where reliability and total cost improve, accounting for repeated context, coordination, and integration; file, line, or acceptance counts are no universal limit. Across tasks check invented or missing dependencies, overlaps, uncovered outcomes, unnecessary prescription while retaining justified contracts and reuse constraints, and bounded review and integration workload. Check the journey's application context and the authority behind consequential product choices. Resolve supported findings before finalizing. If independent review is unavailable, finish preparation and report review as unverified.

Use [tracked work](../skiphow/references/tracked-work.md) to reconcile existing records and save the plan when authorized. Use GitHub Issues within the established audience unless the project uses another authorized tracker. Write engineering artifacts in English. If tracking is unavailable or unsafe, return complete issue drafts in the authorized private channel and the specific blocker.

- Return the specification and issue links or drafts, material decisions, review evidence, unresolved limits, and whether the scope is ready for execution. For ready scope, provide a short execution prompt naming the accepted records, intended destination, scope, and existing authorization boundaries. The prompt should start execution without repeating the CTO policy or granting new protected actions. Keep unfinished preparation and known human prerequisites explicit; a saved plan is not proof of readiness or permission for a protected action. Stop before implementation.

+ Return the specification and issue links or drafts, material decisions, review evidence, unresolved limits, and whether the scope is ready for execution. For ready scope, use this launch shape with the target host's explicit skill invocation syntax. Default to longrun; preserve a different execution workflow the owner already selected.

+ ```text

+ $skiphow-longrun

+ Deliver the accepted implementation scope in <plan-reference> to <agreed-destination>.

+ ```

+ Replace placeholders with concrete accepted records and the agreed destination. Add only otherwise-missing project location, accepted subset, or necessary run constraints, preserving their authorization provenance. Requirements and dependencies stay in the task records; execution policy stays in SkipHow. Grant no new protected actions. Keep unfinished preparation, inaccessible sources, and known human prerequisites explicit as blockers to the affected scope; a saved plan proves neither readiness nor permission. Stop before implementation.