Andie v6.4

Disciplined thinking skill for people who build with AI. Four playbooks — teach, debate, firefight, root-cause — in your LLM.

Runs on Claude Code, Claude.ai, ChatGPT, Codex, Gemini, Perplexity, Manus, and any AI platform. Built by Giggso. MIT License.

You don't ship code without review. Stop shipping decisions without it. Andie runs the playbook the pros actually use — pedagogy, structured debate, incident command, Six Sigma — with a triad of named experts, an adversary in the room, human-in-the-loop gates, and an OODA loop that won't quit until you have a decision. Plan first. Get it right, not just fast.

Who it's for

Solo builders, technical founders, and AI-native operators who make high-stakes calls without a co-founder, a board, or a war room to pressure-test them. Andie is that room — in your terminal.

Why it's different

| | Hands you | You walk away with | |---|---|---| | A virtual eng team (e.g. gstack) | Builders | Code | | An LLM council / panel | Voters | Opinions | | Andie | A discipline | A decision — reviewed, challenged, gated, shipped |

What's New in v6.4

• One hard gate, then value — mode announcement and pre-flight assembly card now arrive in ONE message with ONE "GO". Roster checks, debate-level and Kaizen-cycle continues are non-blocking offers. No more "go" treadmill.
• Implicit GO — uploading a file, pasting a document, or answering a question at any gate counts as consent. Andie incorporates it and proceeds. (Fixes a customer-reported Gemini loop where Andie kept re-asking "rename or start?" after a document upload.)
• Ask-once rule — Andie never repeats a gate; if a reply is ambiguous it states the default and moves on, with an opt-out.
• GATES state ledger — every OODA block tracks which gates passed, so no platform re-asks a passed gate.
• One core, many platforms — all platform files are now generated from core/andie-core.md by build.sh into dist/. New targets: GitHub Copilot, Grok, generic (any LLM). ChatGPT and Perplexity get bootstrap-plus-knowledge-file installs so the full instructions survive their character limits — no more silent truncation, no mojo-losing compaction.

Earlier in v6.3

• Mode splitting — Andie core loads at ~2,100 tokens. Mode-specific instructions load only when that mode is selected. -56% per-message token carry vs v6.2.
• 6 Kaizen methods — Kaizen Cycle, Ishikawa/Fishbone, 5 Whys, DMAIC, Pareto, A3 Thinking. Method selected based on problem shape, not forced.
• Capability routing — capability-map.json maps business needs to specialists. Greenfield: customer picks capabilities. Brownfield: stack detected from project.
• Andie Guru — on-demand Feynman explainer. Say "Guru" or 👍 after any output. 50 words plain English + 100 words Business/Technical/Functional breakdown. Never auto-loaded.
• Domain packs (Tier 2a) — lightweight JSON knowledge files for long-tail tools. Loaded by dynamic-specialist on demand. Agent-frameworks pack (8 tools), local-dev pack (7 tools).
• Open-world tool discovery — tools-landscape gains discover(). Searches MCP registry, validates via guard agents, HITL gates before loading. Trust scoring (HIGH/MEDIUM/LOW).

Earlier in v6.2

• 200-word cap per generation — Andie moves at human pace. One idea per round, fully absorbed before the next.
• Feynman recap after every round — "📌 Here is what we learnt:" — 100–150 word summary combining domain, technical, and data insights.
• HITL tells you what to do — "⏸ APPROVAL NEEDED" with explicit "→ Say go/modify/skip".
• OODA shows progress — mandatory PROGRESS line before every cycle.
• Mode announcement enforced — visible 🎯 mode card with reasoning before work starts.
• RENDER AS per mode — Deep (📘 teacher), Kaizen (🔄 detective), War (🚨 commander), Drama (🎭 writers' room).
• Session goal lock — goal locked at pre-flight, EXIT GATE produces deliverable when done.
• Tiebreaker rules — prevents silent defaulting to Deep.

From v6.1 — all capabilities preserved

Plan mode gate · Multi-platform model routing · Token-optimised · Andie Jr handoff · Skill search · Session memory · Diagram tools

From v5.2 — all capabilities preserved

HITL gates · Specialist Triads · Domain-adaptive questions · Mode previews · OODA as operating rhythm

4 Modes

| Mode | Trigger | What You Get | |---|---|---| | 📘 Deep | Any explanation or learning request | Feynman clarity, layered rounds, devil's advocate | | 🔄 Kaizen | Process improvement, recurring failures | 6 methods: Kaizen Cycle · Ishikawa · 5 Whys · DMAIC · Pareto · A3 | | 🚨 War | Production down, crisis, anything urgent | Incident commander, T+minutes, action owners, war panel | | 🎭 Drama | Decision debate, compare, tradeoff | Named panel argues in first person, 3 debate levels |

Drama is never the default. Max-tier models never used unless you explicitly ask.

Andie Guru — The Explainer

Say "Guru" or 👍 after any output. Guru explains it in plain language:

🧠 GURU — token caching explained

In plain English (50 words max):
Think of caching like a bookmark in a library book...

What this means for you:
📊 Business: Reduces API costs by 40-60% on repeated queries
⚙️ Technical: Add cache headers to prompt prefix, monitor hit rate
🏢 Functional: No workflow changes — it's transparent to end users

One sentence takeaway: Cache your system prompts to cut costs in half.

Never auto-loaded. Never in the way. Always on call.

Specialist Triads

Every domain gets three named experts with personal names from a diverse name pool:

| Domain | 🏢 Functional | ⚙️ Technical | 📊 Data | |---|---|---|---| | ML / AI | ML Product Owner | ML Engineer (training, inference, serving) | ML Data Engineer (features, pipelines) | | Graph Database | Graph Product Owner | Graph Engineer (FalkorDB, Neo4j) | Graph Data Architect | | Workflow | Process Owner | Workflow Engineer (N8N, Temporal, Airflow) | Pipeline Data Engineer | | Oracle (19c/21c/23ai) | Fusion Functional Consultant | Oracle DB/APEX Developer | Oracle Data Specialist | | Salesforce | SF Domain Expert | SF Developer (LWC, APEX, Flow) | SF Data Cloud Architect | | Kubernetes / DevOps | Platform Product Owner | DevOps/SRE (K8s, Docker-compose) | Observability Specialist | | Security | Security Architect | Security Engineer | Security Data Analyst |

Unknown domains trigger dynamic-specialist with domain packs or open-world discovery.

Install

| Platform | File | How | |---|---|---| | Claude Code | skills/andie/SKILL.md | Copy to ~/.claude/skills/andie/ | | Claude.ai | dist/andie.skill | Settings → Skills → Upload | | Gemini · ChatGPT · Copilot · Perplexity · Manus · Grok · anything else | dist/ (one file per platform) | See dist/README.md |

All non-Claude-Code files are generated from core/andie-core.md — edit the core, run ./build.sh, never edit dist/ by hand.

Claude Code — one line:

mkdir -p ~/.claude/skills/andie && curl -fsSL \
  https://raw.githubusercontent.com/giggsoinc/andie/main/skills/andie/SKILL.md \
  -o ~/.claude/skills/andie/SKILL.md

See Install_Anywhere_Agent.md for full platform-by-platform instructions.

Standalone, by design

Andie runs on its own — no other framework required. It works as a self-contained skill on Claude, ChatGPT, Gemini, Perplexity, and Manus, storing its session memory under .andie/.

Larger systems can embed Andie as their thinking layer. Raven v4 does exactly this: it installs Andie (plus andie-jr, andie-frames, andie-guru) alongside its own specialist skills and guard agents. Andie never depends on Raven — the dependency points the other way.

License

MIT — Giggso