Agentic Knowledge Work Method

A folder pattern, four artifacts, and a learning loop for working with AI agents on real, accountable knowledge work.

If you've used Claude or ChatGPT for serious work and noticed your second project doesn't go meaningfully faster than your first — that's not the model. It's the missing scaffolding around the model. This repo gives you that scaffolding.

Read the longer write-up: How I Work With AI Agents — and Why the Method Compounds

What's in here

templates/      drop-in starter files for any new project (LOG, LESSONS_LEARNED, RUBRIC, Research/, README)
skills/         seven Claude Skills that automate the core method
examples/       sanitized excerpt from a real project (a recent guest lecture)
guides/         four short how-to docs covering the trickiest parts
METHOD.md       full method documentation

Quick start (5 minutes)

1. Copy the contents of templates/ into a new project folder.
2. Open README.md and write three lines: who this project is for, what done looks like, and the deadline.
3. Ask your agent to read the project folder and ask you clarifying questions until you both understand the intent.
4. Run the loop (curate research → build rubric → iterate work → log/reflect → mine skills).
5. At project end, run the mine-skills-from-project skill. The output proposes new Claude Skills to add to either the project's .claude/skills/ or your system-wide skills library.

By project five, your skill library is rich enough that the system around the agent — not the agent itself — is the differentiator.

Highly recommended (optional): track your project with git. Each iteration becomes a checkpoint you can revert to if a revision goes wrong. The agent runs every git command for you — you don't need to learn the command line first. New to git? Read guides/05-git-and-version-control.md — it's a friendly explainer aimed at non-technical knowledge workers. The method works without it; the method is significantly better with it.

The method in one paragraph

Treat the agent the way a senior professional treats a high-potential apprentice. Give it clear intent, curated reference material, a quality standard, and a working memory. Capture lessons during the project, not after. At the end, mine those lessons for reusable skills that survive into the next project. The output gets better and faster as the system you've built around the agent compounds.

The five-file folder pattern

Every project, regardless of size or domain:

PROJECT/
├── README.md              what this project is, who it's for, what done looks like
├── LOG.md                 timestamped action log — even meandering steps
├── LESSONS_LEARNED.md     deeper reflection, captured as it happens
├── RUBRIC.md              the quality gate the agent uses for self-evaluation
├── Research/              curated source material, with an index
│   └── README.md
├── work products/         the actual deliverables (slides, docs, code, etc.)
└── .claude/skills/        project-local skills that emerge during the work

The four-stage loop

   1. Clarify intent  →  2. Curate research  →  3. Build rubric
                                                        │
                                                        ▼
       ┌─────────────────────────  4. Iterate work, log, reflect
       │
       │   (loop until rubric passes)
       │
       ▼
       5. Skills mining

Stages 1-3 typically take 60-90 minutes upfront. Stage 4 is the bulk of the project — iterations are assumed, not exceptional. Stage 5 is short and is what makes the method compound.

Skills included

Skill	What it does
init-knowledge-work-project	Scaffold a new project folder with the five-file pattern
clarify-intent	Q&A loop that locks success criteria before any work starts
curate-research-folder	Build an annotated Research/ folder from a topic + a few seed sources
build-project-rubric	Generate a project-specific QA/QC rubric (big-picture + tactical)
log-step	Append a timestamped entry to LOG.md
reflect-lesson	Append a structured entry to LESSONS_LEARNED.md
mine-skills-from-project	At project end, read LOG + LESSONS and propose new reusable skills
init-git-repo (optional)	Initialize git tracking on the project folder with a sensible .gitignore
commit-iteration (optional)	Commit a checkpoint with a useful message after each meaningful iteration round

Each skill is a single SKILL.md file with a description, trigger phrases, and a working prompt. Drop the folder into your ~/.claude/skills/ (system-wide) or PROJECT/.claude/skills/ (project-local).

Guides

• Getting started — first project, step by step
• Writing rubrics that catch issues — the rubric is the quality gate; this is how to make one that works
• Research folder discipline — curated, not exhaustive
• When to promote a skill — project-local vs system-wide
• Git and version control for non-technical knowledge workers — friendly intro for anyone who's never used git; the agent runs every command for you

License

MIT. Use it, fork it, ship it. If you build skills you'd like to share back, PRs welcome.

Author

Built and maintained by Leif Ulstrup — founder of Primehook Technology, MBA faculty at the Kogod School of Business at American University.

This repo is a working artifact, not a product. The method evolves as I run more projects with it. Issues and PRs welcome.