Portfolio · Project 10 · AI in HEOR · October 2026

Turn your skills into an agent

A skill teaches AI how to do one job well. An agent takes a goal and works through it on its own, using several skills, checking its work and stopping when it is done.

This page explains the difference in pictures, shows how to build an agent and why you might want one, and gives you one to download: a protocol author and an independent reviewer that draft and check an observational study protocol together.

Skill or agent? In three pictures

A skill is a recipeIt knows how to do one job, the same way every time. It waits until you ask.

An agent is a workerGive it a goal. It picks the recipes and tools it needs and keeps going until the job is done.

Agents can work as a teamEach one has its own role. They hand the work back and forth until it is right.

With a skillYou drive every step

With an agentYou set the goal, then check the result

Why build an agent

Chain your skills

One goal can run through several skills in order: draft, check, revise. You stop copying output from one chat into the next.

Get a second pair of eyes

A model that checks its own draft tends to go easy on it. A separate reviewer that never saw the drafting reads the work as an outsider would.

Keep going without you

The agent works through each round on its own and reports back. You step in only for the decisions that are yours to make.

Start each job fresh

Each agent gets only the files it needs, not a long chat history, so it is less likely to drift or carry over another task’s assumptions.

When a skill is enough. If you would do the job in one sitting, in one chat, with one skill, keep it a skill. Build an agent when the job has stages, needs an independent check, or repeats often enough that running it by hand gets tedious.

How to build an agent Six steps

Name and description

Who it is, and when Claude should hand it work.

Instructions

How it does the job, in plain English.

Tools

What it may touch: read files, write files, search the web.

Skills

The recipes it can load when the job calls for them.

Its own memory

A fresh workspace that sees only what you pass it.

A stop rule

When it is done, and when to ask a human.

1
Pick a job, not a task
Look for work you already run in stages with more than one skill, or work that needs checking by someone who did not do it. Drafting a protocol and then reviewing it is a good example.
2
Split it into roles
Give each agent one job. Two agents with clear roles work better than one agent asked to do everything. Here there is an author and a reviewer.
3
Write a role card for each
In Claude Code, an agent is a text file in a folder named .claude/agents/. The top of the file holds its name, a description that tells Claude when to use it, and the tools it may use. Below that go its instructions, in plain English. Only the name and description are required. You can also ask Claude Code to write the file for you.
4
Decide what each agent sees
Each agent works in its own context, so pass it files, not conversations. Leaving things out matters as much: the reviewer never sees the author’s notes, so it judges the document the way an outside reader would.
5
Write the stop rules
Say what “done” means (here, no serious issues left open), set a maximum number of rounds, stop if a round makes no progress, and send decisions only a human can make back to you.
6
Test on a case you know, then tighten
Run it on a job you have already done by hand and read every log, the way you would check a new team member’s work. The test of this agent turned up two habits worth fixing, described below.

.claude/agents/protocol-reviewer.md

---
name: protocol-reviewer
description: Independent reviewer
  that critiques an observational
  study protocol and returns a
  READY / NOT READY verdict.
tools: Read, Write, Grep, Skill
---
You are an experienced, independent
HEOR reviewer. You did not write
this protocol and you never edit it;
you only assess it.
…severity scale, checklist, stop rule…

your-project/

.claude/
├── agents/
│   ├── protocol-author.md     the drafter
│   └── protocol-reviewer.md   the checker
└── skills/
    ├── rwe-protocol-writer/   the recipe
    └── protocol-qc-loop/      runs the loop

Claude Code notices new agent files without a restart. The exception is the first time the agents folder is created, when a restart is needed. Agents placed in ~/.claude/agents/ are available in all your projects.

The protocol author–reviewer pair

Two agents and a small coordinator. You hand over a one-page study brief. The author drafts the protocol with the protocol-writing skill. A reviewer, started fresh each round, critiques it. The author revises, and the loop repeats until the reviewer is satisfied.

You

Fill in a one-page brief: question, design, data source, population.

Author

Drafts the protocol, logs every assumption, and stops to ask if a key choice is missing.

Reviewer

Checks the draft against public standards and grades each issue. Never sees the author’s notes

Verdict

Ready, or not ready, with the issues still open.

Not ready? Back to the author, who accepts, rejects or escalates each point. Up to three rounds.

Ready, or three rounds done → you get the final protocolA summary of what changedThe decisions only you can makeEvery draft and log as an audit trail

What a test run showed

Serious issues fell to zero in three rounds

Open major issues, on an illustrative observational study brief. No critical issues were found in any round. The reviewer raised 37 comments in all; 30 were resolved and 2 went to the sponsor to decide.

The independent check earned its place

In round 2 the reviewer caught a problem the author had created while fixing an earlier one: the start date for the main survival endpoint now depended on data that might not be licensed. A model checking its own revision would probably have missed it.

Two habits to fix, now fixed

The author accepted every comment and the protocol grew by 85%. It may now decline suggestions that would not change how the main result is read.
The author filled in a few clinical values from memory, and the reviewer missed them until asked. The reviewer now checks for unsourced values in every round.

Ready to use · version 1.0 · October 2026

Download the agent

The two agents, the coordinator that runs the loop, and a one-page study brief to fill in. The reviewer has a built-in checklist drawn from public standards (STROBE, RECORD, HARPER, TARGET and the ISPOR-ISPE good practices), so it works out of the box. If you have your own review skill, it uses that instead.

Download protocol-author-reviewer.zipZIP · 11 KB · 6 files

Get the skill it uses. Download the protocol-writing skill from Project 02 and unzip it into .claude/skills/ in your project.
Copy the .claude folder from this ZIP into the same project. If that creates the agents folder for the first time, restart Claude Code.
Check it’s there: ask Claude Code “what subagents do you have?” It should list protocol-author and protocol-reviewer.
Run it: fill in the study brief, then type /protocol-qc-loop. You get a one-line update after each round.

Read every file before you run it. Provided as is, without warranty; you remain responsible for any protocol you draft with it, and it needs the same scientific, statistical and compliance review as any other draft. Licensed under CC BY-NC 4.0 for non-commercial use with credit; see the Terms of use.

Built on

Believe.

Design inspired by Spencer Gabor and Dieter Rams’s ten principles of good design.
© 2026 Jigar Rajpura. All rights reserved. No part of this site may be reproduced, distributed or adapted without written permission, except as set out in the Terms of use.

DisclaimerPersonal project, not sponsored by or affiliated with any employer (current or former) or any university I attended; views are my own. Built with AI assistance and checked against Anthropic’s Claude Code documentation as of October 2026, but errors may remain and features change. The test results describe one run on an illustrative brief, not a validation. Full disclaimer