Getting Started
This tutorial installs Accelerator and walks you through your first
full loop — research → plan → implement — on a small change in one of
your own repositories. By the end you will have a research document, a
reviewed plan, and an implemented change, each captured as a file under
meta/ rather than lost in a conversation.
Prerequisites
Section titled “Prerequisites”- Claude Code v2.1.144 or later — earlier versions lack the subagent skill-preload mechanism the plugin relies on.
- macOS or Linux. The shell tooling and the visualiser binary
target both;
jqis required by the session hooks and integrations. - A git (or jj) repository to work in — any project will do.
Install and initialise
Section titled “Install and initialise”-
Add the marketplace and install the plugin. In Claude Code:
/plugin marketplace add atomicinnovation/accelerator/plugin install accelerator@atomic-innovationRestart the session so the plugin’s hooks and skills load.
-
Prepare the repository. From a session in your repo:
/accelerator:initThis creates the
meta/directory tree the skills write to (plans, research, work items, decisions, reviews, …) plus an.accelerator/scaffold for configuration, and gitignores the personal config file. It is idempotent — and optional, since skills also create directories on first use. See The meta/ directory for what lives in each subdirectory. -
(Optional) Add project context. Run
/accelerator:configure createto record your stack, conventions, and build commands in.accelerator/config.md; every skill reads it at invocation time. The configuration cookbook has ready-made recipes.
Your first loop
Section titled “Your first loop”Pick a small, real change — the example below adds input validation to a signup form. The point of the exercise is watching each phase hand off to the next through a file, not the change itself.
-
Research. Ask a question about the code the change will touch:
/accelerator:research-codebase how is the signup form validated today?Claude decomposes the question and spawns parallel subagents — locators to find the relevant files, analysers to read them — then synthesises what they return. You will see something like:
Research complete. Findings written to
meta/research/codebase/2026-07-10-signup-form-validation.md. Key findings: validation is client-side only inSignupForm.tsx:42; no server-side checks exist…Open the document: it contains the findings with file-and-line references. Ask follow-up questions now if anything is unclear — they are appended to the same file.
-
Plan. Start a plan from that research:
/accelerator:create-plan add server-side signup validationThis is a conversation, not a one-shot generation. Expect the skill to push back and ask before it writes anything:
Before drafting the plan I need to resolve a few things the research couldn’t answer:
- Should validation errors reuse the existing client-side message copy, or is new copy acceptable?
- …
Once you have agreed the approach and phase structure, the plan is written to
meta/plans/2026-07-10-add-server-side-signup-validation.mdwithstatus: draft— phases, specific file changes, and success criteria split into automated and manual verification. -
Review the plan (recommended). Before any code exists is the cheapest moment to change course:
/accelerator:review-plan @meta/plans/2026-07-10-add-server-side-signup-validation.mdParallel reviewer agents examine the plan through quality lenses (architecture, correctness, test coverage, code quality) and record a verdict — APPROVE, REVISE, or COMMENT — in
meta/reviews/plans/. Address the findings and re-run until it earns APPROVE. -
Implement. You can do this in a brand-new session — the plan file carries all the context:
/accelerator:implement-plan @meta/plans/2026-07-10-add-server-side-signup-validation.mdThe skill works phase by phase, pausing after each one for your review and ticking success criteria off in the plan file itself:
#### Automated Verification:- [x] Unit tests pass: `npm test`- [x] Type check green: `npm run typecheck`If it is interrupted, re-running resumes from the first unchecked item.
-
Commit and finish. Run
/accelerator:committo turn the session’s changes into well-formed atomic commits, then optionally/accelerator:validate-planto verify the code matches the plan (a pass marks the plandone), and/accelerator:describe-prwhen you raise the PR.
Watch it as a website
Section titled “Watch it as a website”Your meta/ documents are plain markdown, but the plugin ships a
visualiser that renders them as a live local site — plans, research,
and a work-item kanban board:
/accelerator:visualiseWhere next
Section titled “Where next”- Which skill do I need? — map any intent to a skill.
- Development Loop — the ideas behind the three-phase loop you just ran.
- Plan a feature — the loop above as a repeatable how-to, including work items.
- Philosophy — why phases communicate through the filesystem at all.

