Greybeard — Senior Judgment for Coding Assistants

Overview

The big frontier models have judgment. The smaller, cheaper ones have knowledge but no judgment. They’ll happily shuffle your time-series data in cross-validation, run a costless backtest, or call an expensive LLM API for a task a TF-IDF model solves in milliseconds. Not because they don’t know better. Because nobody’s standing behind them saying “stop, that’s the mistake everyone makes here.”

Greybeard is me banking that someone. The old senior engineer who’s seen every mistake and stops you before you make them, written down properly as plain markdown you load into any model’s context. Claude, GPT, Gemini, local models, whatever your assistant is. Load the right file before the task and the model stops making junior mistakes it was always going to make.

Tripwires — the atomic unit

Every file is built around tripwire tables: signal, why it fails, safe move, check. Because honestly, the gap between senior and junior work is rarely the happy path. It’s knowing where the floor is rotten. A real row from the classical-NLP playbook:

Signal	Why it fails	Safe move	Check
First move is an LLM API call for a fixed-label task	10 to 1000x cost, 10x latency, rate limits, for maybe 2 F1 over a fine-tuned encoder	Run TF-IDF + linear as the floor first, then justify each climb	Written cost/latency/F1 line per rung considered

Every check is something a model or a human can actually execute. No vibes.

STAGE MAP — factory to library to your assistant

flowchart TB
    subgraph FACTORY["The Factory"]
        RUBRIC["Rubrics<br/>per asset type"]
        DRAFT["Strong-model drafting<br/>overnight batch runs"]
        ADV["Adversarial reviewer<br/>fresh context, breaks it with evidence"]
    end
    subgraph LIBRARY["The Library — graded, dated"]
        PB["31 Playbooks<br/>tripwires + decision ladders"]
        BP["6 Blueprints<br/>end-to-end buildable specs"]
        SK["11 Skeletons<br/>tested reference code, 160+ passing tests"]
        FM["Failure Maps<br/>where a project shape dies"]
    end
    subgraph LOAD["Load Paths"]
        NPX["npx CLI<br/>Claude · Cursor · Codex · Gemini · Windsurf"]
        PLUGIN["Claude Code plugin<br/>auto-triggering skills"]
        PASTE["plain markdown paste"]
    end
    ASSIST["Your coding assistant<br/>stops making junior mistakes"]
    RUBRIC --> DRAFT
    DRAFT --> ADV
    ADV -->|survived| LIBRARY
    ADV -->|"drew blood"| DRAFT
    LIBRARY --> LOAD
    LOAD --> ASSIST

The library, mapped

The INDEX has one row per shipped asset, and every row carries a load trigger: the phrase pattern a weaker model matches on to know it needs that file. That’s the piece that makes the library work for models and not just humans. The model doesn’t have to know it needs the time-series playbook. The trigger (“forecasting or anomaly detection on time-ordered data”) tells it.

flowchart LR
    subgraph DOMAINS["Domain coverage"]
        MC["Model craft<br/>pretrain · fine-tune · sourcing"]
        ML["ML + CV<br/>classical · deep training · detection"]
        FAM["Task families<br/>time-series · NLP · RL · quant"]
        RET["LLM systems<br/>RAG · retrieval infra · prompts · MCP"]
        AG["Agents<br/>architectures · workflows · evals · safe writes"]
        BE["Backend + applied<br/>services · async · auth · Shopify · scraping · React"]
    end
    METHOD["5 Method playbooks<br/>research · debug · plan · verify · write"]
    TASK["Any real task"]
    METHOD -->|"load alongside"| DOMAINS
    DOMAINS --> TASK

The five method playbooks (how to research, debug, plan, verify, write) load alongside any domain file, and they’re honestly the highest-leverage assets in the whole repo. The intended stack for a task is 2 to 4 files: one method, one domain, plus a skeleton if code is involved.

The factory (this part matters)

None of this is a link dump or a best-practices listicle. Every asset was drafted against a written rubric and then attacked by a fresh-context adversarial reviewer whose only job was to break it with evidence. And it drew blood every single pass. A stale webhook retry policy. A citation transplanted onto the wrong paper. An LR-scaling rule that would have quietly diverged your fine-tunes. What’s in the library survived that, and the assets that haven’t passed yet are marked DRAFT right in the index instead of pretending.

The drafting itself ran as overnight batch jobs, a strong model working through the rubrics while I slept, with the adversarial pass and my own review gating what shipped in the morning. The factory prompts ship in the repo too, so anyone can point a strong model at them and mint playbooks for their own domains.

Graded, not vibes

Every index row carries a status: a dated ✅ means the asset passed its rubric plus a fresh-context adversarial review, and skeletons additionally show their live test counts (14/14, 16/16, 17/17). Fast-moving facts carry a dated currency log with RECON marks, so the file itself tells your model what to re-verify at build time instead of pretending the snapshot is eternal.

Engineering decisions

• Skeletons are pre-tested plumbing. The stuff that fails silently (bounded agent fan-out, leakage-safe CV splits, MinHash dedup, idempotent workers, tenant guards) ships as runnable modules with their own test suites and an ADAPT.md wiring guide. You wire your logic into the marked points and you do not let your model rewrite the concurrency code. That’s the entire point.
• One command, six ecosystems. npx greybeard-playbooks installs into Claude Code (auto-triggering skills), Cursor rules, Codex, Gemini CLI, OpenCode, and Windsurf, with flags to cherry-pick assets. Zero tooling also works, every playbook has skill frontmatter and a plain copy is enough.
• A failure map is its own genre. The quant/trading ML failure map is a step-by-step chart of where that specific project shape dies (leakage, costless backtests, multiple-testing traps) with stop-safely points, distilled from a real system’s real wounds.

Highlights

• 31 playbooks, 6 blueprints, 11 skeletons (160+ passing tests), 1 failure map, all indexed with load triggers
• Every shipped asset survived a fresh-context adversarial review that was paid to kill it, and the ones that haven’t are marked DRAFT honestly
• One-command install across 6 assistant ecosystems, plugin and plain-markdown paths included
• The factory ships with the library, rubrics and prompts included, so the method is reusable, not just the output