agent-pack-core

Evidence-first BDD and research-discipline skills, audit agents, protocols, and rules for Claude Code. Portable, workspace-agnostic, Apache-2.0.

This is not another catalog of "do task X" skills. It is a small, opinionated toolkit for working the way good engineering and good research actually demand: prove it, don't recall it. Every skill cites file:line or command output as evidence; review and audit are first-class, not an afterthought; and the pack improves itself.

What it is

A Claude Code plugin you add to your marketplace and install as a single plugin (all skills become available at once). Drop it into any repo, in any language — the skills detect your stack and read your project's conventions rather than assuming them. There are no workspace paths, no vendored credentials, and no domain assumptions baked in. The benchmark/experiment family reads its domain values from a profile you write (see profiles/_example-profile.template.md); everything else is generic out of the box.

What makes it distinct

Most published skill packs are either Anthropic's reference skills (great building blocks, authoring-focused) or research-content collections. This one is built around four things they do not foreground:

1. Evidence-first, anti-hallucination by default. The anti-hallucination and algorithmic-fidelity protocols and the merge-readiness "trust ledger" make verifiability the default. Skills cite the file and line or the command output that proves a claim, rather than asserting from memory. A claim with no evidence is treated as a rumour, not a fact.

2. Audit agents as first-class citizens. contradiction-hunter, clarity-validator, and ai-research-scientist ship as parallel-dispatch reviewer subagents. Depth review and cross-document contradiction-hunting are part of the toolkit — not just authoring and shipping.

3. Bench / experiment discipline as portable methodology. Paradigm-before-run, never-compare-across-paradigms, pre-registered locked specs (lock-bench-spec), and a four-tier autonomous-loop halt evaluator (check-stop) — research rigor distilled into reusable skills, parameterized by a workspace profile. This is research process, not research content.

4. A self-improvement loop. session-reflect turns observations from a finished session into concrete updates to skills, rules, CI, and ADRs — and the whole pack dogfoods itself (see REVIEW.md). The pack reviews the pack and improves the pack.

Install

Add the marketplace, then install the plugin — all skills become available at once:

/plugin marketplace add izgorodin/agent-pack-core
/plugin install agent-pack-core@agent-pack-core

Or run a local checkout directly without installing:

claude --plugin-dir ../agent-pack-core

Catalog

Skills — BDD / workflow

• describe-task — turn a fuzzy request into a precise behavior statement + deliverables.
• write-failing-tests — write meaningful red tests before any production code.
• make-green — implement the minimum code that turns red tests green; no gold-plating.
• commit-and-pr — run the local quality gate, generate a real commit message from the diff, commit, and open a PR. Refuses to commit if the gate fails.
• address-review — read PR comments, plan per comment (accept / push back / partial), apply, re-run the gate, push — without auto-resolving threads.
• pr-preparation — polish a multi-commit branch for review (squash/reword, PR body from the diff).
• session-reflect — convert a finished session into concrete infrastructure improvements.
• bdd-workflow — the end-to-end 9-step cycle that orchestrates the skills above.
• merge-readiness — GO / NO-GO gate with a trust ledger over the evidence.
• spawn-worker-subagent — fan out work to a worktree-isolated worker subagent.
• goal-launch — launch a goal-directed run (ships launch-windows.ps1 + launch-posix.sh).
• validate-plugin-marketplace — lint a marketplace + plugin manifest pair.
• skill-creator — create, improve, and eval skills (vendored from Anthropic, Apache-2.0).
• onboarding — first-run orientation: what the pack is, the skills, goal-launch safety, and the conventions.
• self-improve — drive the pack's own improvement loop: capture learnings to a backlog, then batch-generate via skill-creator and gate via merge-readiness (never self-merges).
• backport-review — compare a downstream clone against this base and surface universal improvements to port back; run before tagging a clone release. Read-only.

Skills — bench / experiment methodology

• experiment-design — paradigm-before-run discipline.
• experiment-campaign — unattended loop with a budget cap, stop-on-regression, and serialized shared store.
• run-experiment — the pre-flight → background-execute → validate → report → registry → commit spine.
• bench-spec-author — the benchmark-spec authoring interview (class taxonomy, two-axis assembly, verdict logic, anti-Goodhart).
• check-stop — the four-tier autonomous-loop halt evaluator.
• lock-bench-spec — mechanical pre-registration lock + tag.

Agents

• ai-research-scientist — research-grade reviewer subagent.
• clarity-validator — flags ambiguity and underspecification.
• contradiction-hunter — hunts cross-document contradictions.

Protocols

• anti-hallucination — cite evidence; never fabricate structural facts.
• algorithmic-fidelity — numerical/research code must honor its algorithmic guarantees.
• review-and-merge — the depth-review and merge-decision discipline.
• experiment-paradigm-discipline — the three-paradigm model, never-compare-across- paradigms, tier cadence, and registry-row provenance.

Rules

• python-base — baseline Python conventions.
• typescript-base — baseline TypeScript conventions.
• secrets-hygiene — provider-neutral never-commit / never-log / never-paste + leak response.

Profiles

• _example-profile.template.md — neutral stub the experiment skills point at; copy and fill it for your domain.

Companion: Anthropic's reference skills

This pack does not vendor most of Anthropic's reference skills. If you want claude-api, mcp-builder, or pdf, pull them from anthropics/skills and run them alongside this pack — composition, not vendoring. Keeping them upstream means they never rot against their source.

skill-creator is an exception: it is included and vendored in this pack under skills/skill-creator/, attributed to Anthropic (Apache-2.0 — see skills/skill-creator/LICENSE.txt and the NOTICE file at the pack root). The vendored copy is unmodified.

Pinning for consumers

For reproducibility, pin to a tag or commit when you add this pack rather than tracking a moving branch. A skill's behavior can change between versions; pin so your team gets the same toolkit you reviewed. When you bump the pin, run your own review over the diff before adopting it.

License

Apache License 2.0. Copyright 2026 Eduard Izgorodin. See LICENSE.