Skip to content

Workflow Overview

The ACT workflow turns an idea into reviewed requirements, planned implementation tasks, and verified code. Each stage creates focused context for the next stage, with human review points before the agent moves on.

Interview -> Create Spec -> Create Work Items -> Implement

The diagram below shows how each stage turns one artifact into the next:

  1. Task description

  2. act-interview

  3. Rich context and resolved decisions

  4. act-create-spec

  5. Spec and Interview Ledger

  6. act-refine-spec

    Optional quality gate
  7. Spec ready for planning

  8. act-create-issues

  9. Work Items

  10. act-implement

  11. Code changes and verification summary

Use Refine Spec as an optional quality gate for non-trivial work.

Use the plain skills for general work. Use the Flutter/Dart skills when the project or task is Flutter/Dart.

StageGeneralFlutter/Dart
Interview/act-interview/act-interview-flutter
Create Spec/act-create-spec/act-create-spec-flutter
Refine Spec, optional/act-refine-spec/act-refine-spec-flutter
Create Work Items/act-create-issues/act-create-issues-flutter
Implement/act-implement/act-implement-flutter
Skill/act-interview or /act-interview-flutter
InputA task description
OutputRich context and resolved decisions

The interview stage gathers rich context before a single line of code is written. It resolves intent, language, constraints, tradeoffs, and dependencies while changes are still cheap to make.

This matters because late changes are expensive. If the agent misunderstands the goal, uses the wrong project language, or misses a dependency, the mistake can spread through the Spec, Work Items, tests, and implementation. The interview is where ACT slows down enough to make those decisions explicit upfront.

Human checkpoint: Confirm the answers and decisions before creating the Spec.

Skill/act-create-spec or /act-create-spec-flutter
InputCurrent conversation context, usually including interview decisions
OutputA Spec and an Interview Ledger

A Spec is a requirements document that describes what should be built before code changes start. It captures goals, constraints, user flows, acceptance criteria, and important exclusions.

An Interview Ledger is Q/A-first traceability for materially resolved decisions. It is written when the Spec is created. It is not a raw transcript and not a second Spec; it records the important answers that influenced the requirements.

Human checkpoint: Review the Spec line by line. Confirm that it describes the right thing before planning implementation tasks.

Skill/act-refine-spec or /act-refine-spec-flutter
InputA Spec
OutputFindings and suggested Spec updates

Performs an adversarial review pass. It looks for contradictions, missing requirements, weak assumptions, and codebase misalignment before implementation tasks are created.

Use this for complex or high-risk work. For small tasks, you may skip it.

Human checkpoint: Decide which findings to address and update the Spec before creating Work Items.

Skill/act-create-issues or /act-create-issues-flutter
InputA Spec
OutputWork Items

Specs are often too big to be implemented by a single agent in one safe pass. This step decomposes the Spec into independently executable Work Items that contain enough context, acceptance criteria, and dependency information for an agent to execute.

act-create-issues checks unresolved questions, explores the codebase where needed, maps each Work Item back to the Spec and Interview Ledger, and identifies blockers or blocking decisions before implementation starts.

It pauses before writing Work Items so you can review the proposed breakdown, granularity, dependency order, and coverage.

Human checkpoint: Review the Work Items. Confirm that each task is independently executable, correctly ordered, and clear enough to hand to an agent or human.

Skill/act-implement or /act-implement-flutter
InputA single Work Item or Spec
OutputCode changes, verification results, and a summary

Implementation is where the planned work becomes code. act-implement takes a bounded Work Item, checks that blockers are resolved, and keeps the session focused on that unit of work instead of drifting into adjacent tasks.

The end result is code that satisfies the acceptance criteria, fits the agreed scope, and has been checked with the relevant analysis and tests.

Human checkpoint: Review the code, run or inspect the verification results, and smoke test behavior where needed.

Not every task needs the full pipeline:

  • Small task: Create a Spec, then run /act-implement directly on it
  • Simple feature: Interview, create a Spec, create Work Items, then implement one Work Item at a time
  • Complex feature: Include Refine Spec before creating Work Items
  • Flutter/Dart feature: Use the -flutter workflow path instead of expecting Flutter behavior to activate inside the plain skills