Introducing AkiDevRule & akiflow: A New Standard in AI Agent Council Orchestration
AkiDevRule and the akiflow skill officially launch, introducing a standardized rule corpus and Agent Council mechanism to eliminate context flooding, role collapse, and self-approval in software development.
AkiDevRule and the akiflow skill officially release today, delivering a standardized rule corpus and Agent Council operating protocol to eliminate hallucinations, self-congratulatory code approval, and context flooding during complex software engineering tasks.
When building large software systems, a single-thread LLM context quickly reaches architectural limits. Upgrading infrastructure or broadening prompts alone cannot resolve these structural failure modes.
Why a Single AI Agent Cannot Solve Complex Software Tasks Alone
Through real-world software engineering, single-thread LLM contexts experience three primary structural failure modes:
Context Flooding: Cramming codebase context, task plans, diffs, UX requirements, and review feedback into a single prompt window severely reduces recall accuracy and increases hallucination rates.
Role Collapse: An AI playing Architect, Coder, UX Designer, and Reviewer simultaneously applies a single compromised standard of correctness across conflicting domain constraints.
[CAUTION] Self-Approval (Sycophancy): An agent generating code cannot objectively review its own work. Forking a reviewer subagent inherits the lead’s biased self-justification chain.
AkiDevRule: Single Source of Truth (SSoT) with topic.A1 Addressing
AkiDevRule resolves AI instruction drift by establishing a layered Single Source of Truth (SSoT) rule corpus:
The Core Tier holds invariant constraints loaded via @-import: index.md, RULE-agent-behavior.md, RULE-coding.md, RULE-pattern-core.md. The Contextual Tier holds domain-specific rules (e.g. RULE-docs.md, RULE-biz.md, RULE-release.md, RULE-seo.md, RULE-ui-pattern.md). The Analytical Tier contains deep-thinking frameworks (METHOD-deep-think.md, METHOD-flow-audit.md, METHOD-ux-psych.md).
Every rule uses precise topic.A1 addressing (e.g., coding.B3 for external action completeness; agent.B5 for read-only audits), allowing AI agents to retrieve exact rules instantly without hallucinating patterns.
How the akiflow Skill Operates the AI Agent Council
akiflow runs on two foundational laws rather than a fixed activation gate: R1 — Anchor, the owner's words are immutable and the final test, the lead is never allowed to restate them; and R2 — Justification, every mechanism — seat, check, script — defaults OFF and turns on only when this specific run produces a real reason.
Two shapes, one frame: council when there is genuinely something to arbitrate — decomposable, with more than one independent standard of "correct"; dispatch when the work is merely large, not adversarial, sharing the same anchor and checklist but split into lanes with exclusive file ownership instead of items with a challenger.
Five fixed agent definitions: aki-hands does pure retrieval, judgment forbidden; aki-judge judges against exactly one standard named at spawn time; aki-conduct audits the process itself via scythe.py; aki-challenger attacks a finished result from a clean context; aki-maker is the only agent permitted to write code, and only in execute mode.
Three gating scripts: council_open.py pins the owner's words and refuses to open a room until the checklist is cut; council_read.py locates before it reads, instead of scanning the whole room; council_verify.py blocks an item from closing if it's missing a [RULES] receipt, missing a FACT/ASSUMPTION tag, or if any REQ is left uncovered by every item's covers.
Real-world Applications of AkiDevRule and akiflow in the AkiTao Ecosystem
AkiDevRule and akiflow have been battle-tested in Nuxt 4 / Cloudflare upgrades, cross-device UI/UX testing, and data consistency audits across the AkiTao ecosystem.
Results achieved: Over 80% reduction in silent self-approval bugs, absolute architectural stability, and 60-70% savings in token expenditure via prompt caching and proper decomposition.
As of version 2.5.0 (2026-08-19), AkiDevRule has grown to ten skills. The tenth, /akiship, turns the daily release ritual — lint, diff-scoped hygiene cleanup, doc sync, grouped commits, deciding whether to mint a version or defer it, generating artifacts — into a single invocation, orchestrating RULE-release.md B7/B8 rather than restating it.
How to Experience and Adopt AkiDevRule & akiflow
Quick & Safe Integration: AkiDevRule and akiflow integrate seamlessly into Claude Code, Antigravity CLI, and existing AI Agent environments with a single command.
The entire project is released as open-source software under the MIT license, with full documentation and source code maintained on GitHub.