SuperAPI AI Directory

task-preparation

Coding & development

View repository

Updated 10-15 10:05

About this skill

Task Preparation

Purpose

Turn one selected roadmap task into implementation-ready task-local docs. This is the spec stage.

Use when

  • one concrete task is selected or clearly next from confirmed roadmap state
  • the next work is task-local clarification, spec.md, optional plan.md, and execution handoff
  • roadmap structure is stable enough that implementation can be prepared without reshaping milestones or tasks

Do not use when

  • milestone confirmation, task selection, or dependency order is still ambiguous
  • the current work is still roadmap decomposition or planning alignment
  • the current work is implementation, readiness, verification, or branch closing

Read first

  • selected task task.md
  • relevant milestone or module index.md
  • relevant docs/architecture/ constraints
  • supporting docs/context/ notes when needed
  • the current user prompt

Owns

  • task selectability checks
  • task-local clarification framing for grilling
  • spec.md
  • optional plan.md
  • review pauses for task-local docs
  • default reviewed-doc follow-up
  • execution handoff context

Must not own

  • milestone reshaping or new roadmap decomposition
  • execution isolation, implementation, verification, or branch closing
  • leaving blocking task-boundary ambiguity hidden inside a vague spec
  • using plan.md to introduce new scope or rewrite the approved spec

Entry checks

  • Enter only when all are true:
    • the milestone has Roadmap confirmed: yes
    • previous milestone closure is resolved when crossing milestones
    • task status is planned or in_progress
    • dependencies are satisfied or explicitly waived
    • no unresolved hard gate blocks progress
    • the user selected the task, or docs clearly identify it as next
  • If any of those conditions are unclear, route back to milestone-planning.
  • Use the external grilling skill for task-local interview loops. If it is unavailable, ask the user to install it instead of recreating its full protocol here.
  • Always write spec.md. Do not skip it.
  • Default to no plan.md. Create one only when a real plan trigger is present.

Clarification rules

  • Run a grilling session before drafting spec.md, even when the task looks obvious.
  • Use code and docs to resolve present-state facts first.
  • Do not treat current implementation or existing docs as authority for future intent, scope, acceptance, or non-goal decisions.
  • Ask the user whenever the unresolved branch is about what should change, what should stay out of scope, or what future behavior the task is meant to produce.
  • Let grilling run the interview protocol while this skill preserves task-local boundaries.
  • Treat scope, acceptance, non-goals, migration ownership, rollout behavior, proof expectations, and allowed minimal implementation as blocking when they materially affect the task-local design.
  • Do not draft from assumed defaults while blocking branches remain unresolved.

Default flow

  1. Confirm the selected task is genuinely selectable from confirmed roadmap state.
  2. If the task is still planned, move it to in_progress when active task-local work begins.
  3. Run grilling with task-local constraints until blocking branches are resolved.
  4. Draft the full spec.md in one pass.
  5. Stop for spec review.
  6. If a plan trigger is present after spec review, write plan.md and stop for plan review.
  7. Run the default docs follow-up.
  8. Hand off to execution only after the docs follow-up outcome is explicit.

Plan triggers

  • 3 or more major files or modules must change and modification order affects correctness
  • database schema, migration, backfill, or persisted-format changes
  • public API, CLI, config format, plugin interface, or task-doc compatibility boundary
  • cross-module coordination where responsibilities or sequence are non-obvious
  • phased rollout, feature flag, dual-read, dual-write, or migration transition behavior
  • non-obvious verification order
  • multiple implementation slices remain inside one task and should not be split at roadmap level
  • exploratory spike or risk-reduction work is required before implementation edits

Do not create plan.md for a small single-capability task with straightforward implementation order. If plan-trigger status is uncertain, name the suspected trigger and ask before writing the plan.

Boundary rules

  • Keep 1 task -> 1 spec. If the task appears to need multiple specs, the task boundary is wrong.
  • plan.md supplements the approved spec.md; it does not restate the spec.
  • Do not write follow-up work into the current plan.md unless it is already in the current task scope.
  • Future work discovered during preparation belongs in the current milestone index.md under Handoff Notes by default, not in plan.md.
  • Compact specs are fine; vague specs are not.
  • Make boundaries, non-goals, likely first edit surface, and proof expectations explicit enough for a fresh conversation or another agent to continue.

When to route back

Route back to milestone-planning instead of forcing the current spec when any of these become true:

  • the task needs two mutually exclusive scope choices to make sense
  • the task appears to need two separate specs
  • acceptance actually describes two independently reviewable or independently shippable capabilities
  • a missing dependency is really a roadmap-order problem rather than a task-local design question

Task status rules

  • Entering active task-local preparation should move the task from planned to in_progress if it has not already changed.
  • Keep the task at in_progress while clarification, spec.md, optional plan.md, review pauses, or docs follow-up are still active.
  • Use blocked when task-local preparation cannot continue because a real blocking question, external dependency, missing prerequisite, or explicit user hold prevents progress.
  • Move a blocked task back to in_progress once the blocking condition is resolved and active work resumes.
  • Do not mark the task completed during preparation; completion belongs to execution after implementation and branch closing are done.

Review and default follow-up

  • Stop at a review pause after spec.md.
  • Stop at a second review pause after plan.md when a plan exists.
  • Treat any clear forward-motion message after a review pause as permission for routine follow-up.
  • At a spec.md review pause, that same forward-motion message counts as approval of the reviewed spec unless the user asks for changes instead.
  • Default follow-up after the final task-local review is:
    • commit reviewed task-local docs
    • unless the user explicitly asks to leave them uncommitted
  • If reviewed docs stay uncommitted, say that it is an explicit exception, give the reason, and list the affected files.
  • The stage is not complete until the docs follow-up outcome is resolved and reported.

Hard gates

  • Do not use continue to bypass unresolved task-boundary ambiguity.
  • Do not hand off to execution while a blocking Undecided item still affects spec.md, plan.md, or the reviewed-doc follow-up.
  • Do not let reviewed-doc approval imply permission for destructive execution actions; those belong to the execution stage.

Output

  • Task
  • Readiness check
  • Blocking questions
  • Spec result
  • Plan needed? why?
  • Review stop
  • Docs follow-up outcome
  • Execution handoff

Stop point

  • spec review pause
  • plan review pause
  • or a direct blocking question about task-local ambiguity or task boundary failure

Handoff

  • Decided
  • Undecided
  • Next skill
  • Stop point

References

  • Generate spec.md in the repository documentation language when a language rule exists; otherwise use the current user language.
  • Template section names are structural hints, not fixed English text. Localize headings, labels, and field names to the selected documentation language.
  • Use references/spec-template.md for task-local spec.md.
  • Use references/plan-template.md for task-local plan.md.