Why Split Rule Loading Into Two Tiers?
A skill only runs when the model itself decides to invoke it — meaning any rule that lives only inside a skill is "best-effort loaded" and can be skipped if the model never thinks to call that skill. AkiDevRule addresses this by clearly separating two different loading mechanisms instead of treating every rule the same way.
Core Tier — Hard-Loaded by the Harness, No Model Decision Involved
Four files — index.md, RULE-agent-behavior.md, RULE-coding.md, and RULE-pattern-core.md — are @-imported directly into ~/.claude/CLAUDE.md, which the Claude Code harness reads mechanically the moment a session starts, with no model choice to invoke any skill required. These are the only rules that apply to every task, even ones with no code involved. The cost is that these four files occupy context in every session, including sessions that don't need them — a deliberate price for a hard guarantee.
RULE-agent-behavior.md carries rules like: communicate before acting (a question is not a request to execute), dense reports over verbose ones, decision boundaries (when to ask before acting), and memory discipline. RULE-coding.md carries code philosophy: single-maintainer-friendly, MVP-first, DRY without premature abstraction, comments as a last resort. RULE-pattern-core.md carries 8 universal architecture laws: Single Source of Truth, Rule of Three, Single Responsibility, Open/Closed, composition, module boundaries, naming by role, and "one natural flow instead of stacked guards".
Contextual Tier — the akirule Skill Routes on Signal
Every other rule — RULE-docs.md, RULE-content-write.md, RULE-stack-akiNuxtCf.md, RULE-stack-tauri.md, RULE-ui-pattern.md, RULE-seo.md, RULE-release.md, RULE-db-design.md, RULE-biz.md, plus the METHOD-*.md files (deep-think, flow-audit, zero-trust-audit, ux-psych, proportionality, subtraction-audit) — only loads when the akirule skill is invoked and a signal matches (the file type being edited, the task's topic). Signal sensitivity is deliberately set high: better to over-load a few tokens than miss one and misbehave. Say "load all rules" (or the Vietnamese "nạp full") to load the entire corpus at once regardless of signal.
Three files — RULE-seo.md, RULE-release.md, and RULE-stack-akiNuxtCf.md — mix universal rules with content specific to the AkiNuxtCf ecosystem (the usePageSeo API, the releases.json schema, canonical component naming...). That ecosystem-specific content is isolated into the last group of each file, flagged ⟨Aki⟩ — it stays in the public repo and still auto-loads, but the flag marks exactly what a stripped public-only release would drop.
Addressing Rules: topic.A1
Every rule/method file is organized into A/B/C groups with numbered items inside, e.g. coding.B2 (changing existing code), stack.C1 (canonical component naming) — letting any single rule be cited precisely without renaming a file or touching the routing mechanism. topic is simply the filename with its RULE-/METHOD- prefix dropped.
10 Skills on the Open SKILL.md Standard
SKILL.md is an open standard multiple CLIs can read, so AkiDevRule's 10 skills (akirule, akiflow, akithink, akihtmlreport, akihelp, akigitcommit, akilint, akiship, aki-article-writer, akidevsync-notes) are synced unmodified to Antigravity, Codex CLI, Kiro CLI, and Grok CLI, not just Claude Code. Among these, akiflow is a lead-coordinated agent council for work needing more than one kind of judgment, while akiship runs the full release ritual end to end under an autonomy contract with clear stopping points (public-history ambiguity, unclassifiable work, or a design contradiction).
5 Agents — Write Access Limited by Mechanism, Not a Promise
The 5 agent files (aki-hands, aki-judge, aki-conduct, aki-challenger, aki-maker) are written in Claude Code's own ~/.claude/agents/ format, so three properties that used to be prose become mechanical: read-only access is enforced by the tools: field simply omitting Edit/Write, the model tier is declared explicitly in model: instead of being improvised mid-run, and an undefined seat simply cannot be convened. Of these 5 agents, only aki-maker is granted Edit/Write — the other four, including aki-conduct (the agent that checks whether a rule was actually followed), are read-only.
Because an agent-definition format currently has only one consuming CLI (Claude Code), these 5 agents sync to Claude Code only, unlike the skill set which syncs across CLIs. See the product overview in What Is AkiDevRule? or the install steps in Installing AkiDevRule for Claude Code. Full source and technical docs at github.com/lacvietanh/akidevrule.