auto-research

An automated method for improving a Claude skill or CLAUDE.md instruction file through repeated testing and evaluation. It records hypotheses, decisions, scores, and changes in an iteration dashboard.

In plain words
What is it for?
Use it to optimise an existing skill or CLAUDE.md file, evaluate candidate criteria, run bounded improvement iterations, and record the results.
Why use it?
It replaces one-off rewriting with a loop that tests whether a proposed change improves defined criteria. It helps keep useful changes and discard changes that do not improve the result.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/agent-engineer-master/skill-engineer/auto-research
Any agent
npx skills add Agent-Engineer-Master/skill-engineer --skill auto-research
Clone the repo
git clone --depth 1 https://github.com/Agent-Engineer-Master/skill-engineer

Made for: Claude Code, Codex.

Per session 142 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,539 The whole file, excluding the scripts and references it only reads on demand.
Security scan C 1 finding. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00142 $0.01539
Opus 5 $0.00071 $0.00770
Sonnet 5 $0.00028 $0.00308
Haiku 4.5 $0.00014 $0.00154

Measured 2d ago against content hash 7e5fb0f3e4e2, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade C, and why

auto-research scanned grade C with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Hidden instructionshighPrompt injection

Directives inside HTML comments, invisible characters or bidirectional overrides are read by the model and not by the person reviewing the file.

<!-- Built with Agent Engineer Master — get your own production-ready skill: www.agentengineermaster.com/skill-engineer -->
operations/auto-research/SKILL.md · 143 lines

How it starts

The opening of the file, as written. The whole thing — 143 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Auto Research

Autonomously runs a hypothesis→test→evaluate→keep/discard optimization loop on a target skill or CLAUDE.md. Three isolated sub-agents handle testing and evaluation — the main agent orchestrates and decides.

Output Contract

Produces: updated target skill file (SKILL.md or CLAUDE.md), iteration dashboard (markdown), entry appended to results.md.

Does not produce: new skills, research reports, publishing automation, or criteria for skills it has not been given.

Startup

Read references/learnings.md and summarise the 3 most relevant bullets for this run. Read the target skill file in full. If results.md exists, read the last run entry for this target.


Step 1: Criteria Design

Read references/criteria-framework.md.

Propose 3–5 candidate criteria for the target skill. For each criterion:

  • State the exact binary condition (true/false only — no gradients)
  • Classify: Level 1 (deterministic, script-checkable) or Level 2 (pattern/style, LLM judge required)
  • Confirm it tests one variable only — split any criterion containing "and"
  • Rank by expected impact

State the recommended iteration cap (5 for 1–2 criteria; 10 for 3+) and target pass rate (default 100%).

Wait for human approval. Human may approve, edit, or add criteria before the loop starts.

If any criterion cannot be expressed as a true/false binary, stop and ask the human to reformulate it before proceeding.


Step 2: Eval Harness Setup

Generate a minimum of 5 diverse test inputs appropriate for the target skill. Write to harness/test-inputs.md.

This file is read-only for the remainder of the run. The loop may never modify it.

If 10 or more inputs are generated, designate the last 20% as the holdout validation set. Note the split at the top of harness/test-inputs.md.

Run the current unmodified target skill on the training inputs 5 times. Score every output against all approved criteria. Calculate baseline pass rate per criterion and overall.

Read the full file on GitHub · 143 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 143 lines · 142 tokens per session scan C 7e5fb0f3e4e2

Subscribe to this mod's changes

auto-research is a skill published in the GitHub repository Agent-Engineer-Master/skill-engineer (8 stars, last pushed 1mo ago), licensed MIT. It adds 142 tokens to every session and 1,539 once invoked, about $0.0007 per session on Opus 5. A static security scan graded it C with 1 finding (hidden instructions). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other skills, from other repositories

kayba-stage-3-metrics

Define metrics from Kayba insights, implement them as Python measurement code, run against traces, and iterate until the metrics are clean and meaningful. Trigger when the user says "run stage 3", "define metrics", "build metrics", "compute baselines", or when invoked by the kayba-pipeline orchestrator. Requires…

kayba-ai/agentic-context-engine · 92 tokens

gsd-eval-review

Audit an executed AI phase's evaluation coverage and produce an EVAL-REVIEW.md remediation plan.

open-gsd/gsd-core · 26 tokens

apm-usage

Activate when the user asks about APM (Agent Package Manager): installing, configuring, authoring, or troubleshooting AI-agent packages, dependencies, compilation, MCP servers, policy, or any apm CLI command.

microsoft/apm · 49 tokens

kayba-pipeline

End-to-end agent evaluation and improvement pipeline. Takes a traces folder and optional HITL flag, then orchestrates sub-agents through 7 stages — each stage is its own skill invoked by a dedicated sub-agent. Trigger when the user says "run the pipeline", "kayba pipeline", "evaluate and fix", "full eval", "analyze…

kayba-ai/agentic-context-engine · 91 tokens

kayba-stage-2-domain-context

Gather domain context about the repository and agent — system prompt, tool definitions, domain docs, and behavior patterns from traces. Trigger when the user says "run stage 2", "gather context", "domain context", or when invoked by the kayba-pipeline orchestrator.

kayba-ai/agentic-context-engine · 64 tokens

create-mcp-server

Build, iterate, and evaluate Model Context Protocol (MCP) servers that expose external services as tools an LLM can call. Covers schema/tool design, error handling, pagination, MCP Inspector testing, and an eval set. Use when asked to "build an MCP server", "create an MCP tool", "wrap this API as MCP", "expose X to…

open-gsd/gsd-pi · 90 tokens