aims

aims

AI Manager System

📄 Site: https://eliezeravihail.github.io/aims/

Lean code-development discipline for Claude Code. Two slash commands, project-local hooks that inform but never block, idempotent bootstrap. No multi-agent pipeline, no orchestration overhead — just the discipline that makes single-dispatch sessions reliable on Opus / Sonnet baselines.

What this is for (and what it isn't)

The point of aims is not to make the agent smarter or more correct. The model's reasoning capability is whatever it is — aims doesn't change it.

What aims actually does:

• Keeps the human side of the work organized. Plans on disk, decisions in ADRs, a router that asks before edits — a workflow you can fall into without having to remember ceremony every time.
• Lets the agent know the project better. CLAUDE.md, the ADR log, and plans on disk all become durable context that survives session compaction and crosses sessions. The Claude session that picks up your work tomorrow has access to what was decided yesterday and why.

What aims explicitly doesn't try to do:

• It doesn't change how the model reasons.
• It doesn't turn wrong answers into right ones.
• It doesn't substitute for tests, domain knowledge, or careful prompts.
• It doesn't add an "intelligence layer" via routers, validators, or multi-agent orchestration. (That was the previous design; see ADR-0002 for why we dropped it.)

If the agent is making bad calls, aims will not fix that — better tests, clearer requirements, or a different model will. aims addresses a different problem: the human-side cost of remembering what was decided and why, session after session.

What you get

Planning is a project behavior, not a command to remember (ADR-0022). For a non-trivial change, the assistant runs read-only discovery → writes a Status: draft plan to docs/plans/ → asks for approval → implements → inline close-out. The prompt-submit hook describes this convention factually for every actionable prompt.

Two slash commands exist as optional shortcuts (see ADR-0010, ADR-0022):

Command	What it does
/plan <task>	Dispatches Phase 1-2 to an Opus subagent (read-only discovery + draft write); main session resumes for approval / implementation / close-out. Use when the session model is not Opus.
/install-on <path>	Bootstrap (or idempotently re-install) ADRs, hooks, memory tree, CLAUDE.md.

Everything that used to be its own command now happens inline, with no command to remember:

• Plan close-out (verify steps, run ## Verification, decide ADRs, mark the plan completed, consolidate memory) runs at the end of the implementation session, nudged by the Stop hook when an in-progress plan exists.
• ADRs are auto-decided per change: created when it's a clear architectural commitment, skipped for bug/refactor/doc/test/mechanical work, asked only when borderline. They always start proposed.
• Memory bootstrap runs at the end of /install-on; maintenance after that is the automatic marker + consolidation loop (ADR-0007 / ADR-0009).
• Mechanical edits and notes are just ordinary edits — do the work.

/plan does NOT switch the main session model — only the Phase 1-2 subagent runs on Opus (ADR-0022). Implementation and close-out run on whatever the main session is on.

Hooks (per-project, installed by /install-on)