Working the pipeline

Before you start

This page is the how — a step-by-step guide to driving a real piece of work through the AID methodology from discovery to deployment.

Before running the pipeline, make sure AID is installed and added to your project. See the Installation guide for the one-liner bootstrap and the aid add <tool> step.

For the why behind each phase — the mental model, the philosophy, the full methodology explanation — read the Methodology concept page. This guide does not duplicate that depth; it points you at the right command and artifact for each step.

The pipeline at a glance

flowchart LR
    C(["/aid-config<br/>Step 0"])
    D(["/aid-discover<br/>Phase 1"])
    I(["/aid-interview<br/>Phase 2"])
    T{"TRIAGE<br/>full or lite?"}
    S(["/aid-specify<br/>Phase 3"])
    P(["/aid-plan<br/>Phase 4"])
    DT(["/aid-detail<br/>Phase 5"])
    E(["/aid-execute<br/>Phase 6"])
    DP(["/aid-deploy<br/>Deliver"])
    M(["/aid-monitor<br/>Deliver"])

    C --> D
    C --> I
    D --> I
    I --> T
    T -- "full path" --> S
    S --> P
    P --> DT
    DT --> E
    T -- "lite path<br/>skips Specify/Plan/Detail" --> E
    E --> DP
    DP --> M
    M -->|"findings loop back"| I

The diagram shows the canonical order. Greenfield projects skip Phase 1 (Discover) and go directly from Configure to Interview. The Deliver skills (deploy + monitor) are optional and loop findings back into Interview.

Step 0 — Configure (/aid-config)

Run /aid-config once before starting any new piece of work. It scaffolds or inspects .aid/settings.yml and lets you set the project name, review grade threshold, and other workspace-level options.

1. Invoke the skill — type /aid-config in your AI tool.

2. Review the generated settings file — the skill creates or updates .aid/settings.yml. Confirm the project.name and review.grade fields.

3. Commit the settings file — .aid/settings.yml is checked in. All agents read it; keep it accurate.

Tip

Re-run /aid-config any time you want to inspect or change workspace settings. It is idempotent — running it on an existing project does not reset your configuration.

The six phases

Each phase has a corresponding /aid-* skill. Invoke the skill in your AI tool; the skill runs its state machine and produces the named artifact.

Phase 1 — Discover (/aid-discover)

Brownfield projects only. The Discover skill analyses an existing repository and populates the Knowledge Base with what it finds (architecture, conventions, existing patterns). State machine: GENERATE → REVIEW → … → DONE.

Artifact produced: .aid/knowledge/ entries describing the existing codebase.

Note

Greenfield projects skip this phase. If you are starting a brand-new repository with no existing code, go directly to Phase 2 (Interview).

Skills reference Full roster of AID skills with descriptions and argument hints.

Phase 2 — Interview (/aid-interview)

The Interview skill gathers requirements conversationally. It asks you structured questions about your goals, constraints, and scope, then writes REQUIREMENTS.md and decomposes the requirements into feature files under .aid/<work>/features/.

Artifact produced: REQUIREMENTS.md + features/feature-NNN-*.md stubs.

Skills reference Full roster of AID skills with descriptions and argument hints.

Phase 3 — Specify (/aid-specify)

The Specify skill writes a technical specification for one feature at a time. Point it at a feature file; it reads REQUIREMENTS.md, the Knowledge Base, and the feature stub, then appends a SPEC.md section with design decisions and an acceptance-criteria mapping.

Artifact produced: SPEC.md appended to features/feature-NNN-*.md.

Run once per feature — repeat this phase for each feature before moving to Plan.

Skills reference Full roster of AID skills with descriptions and argument hints.

Phase 4 — Plan (/aid-plan)

The Plan skill reads all your feature specs and sequences them into a delivery roadmap. It decides which features go into which delivery, respecting dependencies, and writes PLAN.md with the strategy.

Artifact produced: PLAN.md — the delivery sequence and dependency graph.

Skills reference Full roster of AID skills with descriptions and argument hints.

Phase 5 — Detail (/aid-detail)

The Detail skill breaks a delivery into typed tasks and produces an execution graph. Each task gets a task-NNN.md file with one of the eight task types (RESEARCH, DESIGN, IMPLEMENT, TEST, DOCUMENT, MIGRATE, REFACTOR, CONFIGURE), acceptance criteria, and dependencies.

Artifact produced: task-NNN.md files under .aid/<work>/tasks/ + an execution graph.

Run once per delivery — repeat for each delivery in your plan.

Skills reference Full roster of AID skills with descriptions and argument hints.

Phase 6 — Execute (/aid-execute)

The Execute skill picks up a typed task and implements it on a per-delivery branch, with a built-in review loop. The Developer agent writes the code; the Reviewer agent grades it against the task’s acceptance criteria; failing grades loop back for revision.

Artifact produced: committed code on a delivery branch, graded and merged.

Run once per task — repeat for each task in the execution graph, respecting the dependency order.

Tip

The Execute skill tracks the current delivery branch automatically. Run tasks in dependency order; the skill warns you if a dependency is not yet complete.

Skills reference Full roster of AID skills with descriptions and argument hints.

Delivering: deploy and monitor (optional)

After all Execute tasks are complete, the optional Deliver skills package and observe the work.

Deploy (/aid-deploy)

The Deploy skill packages completed deliveries into a release — tarballs, changelogs, and registry publishes. It reads the delivery branch state and the VERSION file, and invokes the release pipeline.

Monitor (/aid-monitor)

The Monitor skill observes the deployed system, classifies incoming findings (bug / enhancement / question), and routes actionable items back into /aid-interview as new requirements. This closes the feedback loop.

Note

Neither Deliver skill is required to use AID. Many teams run the six core phases and handle deployment through their existing CI/CD pipeline without /aid-deploy or /aid-monitor.

Lite vs full path

Not every project needs all six phases. For small, well-scoped work, /aid-interview’s TRIAGE routes down the lite path — Interview → Execute, skipping Specify, Plan, and Detail — emitting a work-root SPEC.md with tasks directly. The Methodology page explains when TRIAGE chooses the lite path and when the full pipeline pays off.

The Installation guide also covers the quick-start flow for new adopters who want to get something working before reading the full pipeline.

Next steps

CLI reference All AID subcommands, flags, and exit codes.

Skills reference Full roster of AID skills with descriptions and argument hints.

Agents reference The specialist agents that AID dispatches during Execute.

Methodology The why behind the pipeline — mental model and design decisions.

Maintainer guide Cut a release and regenerate host-tool install trees.