Hoard What You Know How To Do
Keep a searchable backlog of verified, working experiments so agents can recombine them into new solutions.
- Difficulty
- Easy
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 92%
A career-long value-building practice, supercharged by AI. You accumulate a large backlog of techniques and tools you have actually tried and verified work, so that when a new problem appears you can recombine past solutions to solve it — possibly as the only person in the world who has tried that particular combination. AI makes building the backlog nearly free and makes it directly consumable: agents can search an entire drive of prior work to assemble a new answer.
Origin
Framed by Simon Willison as lifelong career advice that also happens to make AI agents more effective. He maintains public GitHub repos (simonw/tools, ~193 client-side tools; simonw/research) plus private repos and ~10,000 Apple Notes.
Core principles
- 01Value comes from being able to combine techniques you alone have touched to solve a new problem
- 02AI makes the cost of trying and recording a new technique nearly zero
- 03Coding-agent research (code actually written and run) is actionable; unverified deep-research reports are 'LLM vomit'
- 04Default to public — it is easier to find later and builds credibility
- 05Agents can search a whole hard drive to find just the examples they need, so context length no longer limits reuse
How to run it
- 1
Pick a note store you trust and won't lose
Choose a durable, backed-up home for your backlog — a Dropbox-synced folder, or GitHub (free private repos, replicated to three continents).
Pro tip GitHub doubles as a backup system and, if public, builds your credibility as a builder.
- 2
Capture small tools that each embody one idea
Whenever you build or have an agent build a little HTML/JavaScript tool, save it. Each one records that a particular thing is possible to do.
Pro tip Willison's simonw/tools repo holds ~193 of these; many are trivial, but each captures a reusable capability.
- 3
Run verified coding-agent research and save the output
Point a coding agent (often on your phone) at a new library or question: download it, try it against a real problem, run the code, and write a markdown report. Save that report.
Pro tip Because the agent actually ran the code and plotted the result, the output is verified and actionable — not an unverified web summary.
Watch out Do not confuse this with traditional deep-research reports; a repo of unverified reports is 'very little value to anyone.'
- 4
Feed the backlog back into agents by reference
Tell the agent to consult specific past artifacts and combine them — e.g. 'check out simonw/research, look at the WebAssembly and Rust ones, and use that to solve this new task.'
Pro tip Paste in URLs to two or three prior tools and say 'read the source and solve this new problem' — agents are exceptional at recombining supplied context.
In the wild
Willison had earlier built two separate browser tools — one using Mozilla's PDF library to render PDFs, one using Tesseract for in-browser OCR. Wanting to OCR PDF pages, he handed both codebases to Claude and asked it to build a new tool that opens a PDF and OCRs every page.
→ The agent combined the two hoarded tools into a working new capability immediately.
Common mistakes
Hoarding unverified research reports
A backlog of deep-research summaries the agent never actually ran is nearly worthless; the value comes only from experiments where code was written and executed.
Keeping it in a store you might lose
If the backlog lives somewhere untrustworthy or un-backed-up, you lose the compounding value; use a durable, replicated store like GitHub.
Is it for you?
Best for
Individual engineers and technical builders who want their experience to compound and feed directly into agent workflows
Not ideal for
People unwilling to maintain a note/repo system consistently, for whom the backlog will go stale
From the transcript
“the way you build value as a software engineer or pretty much any other profession is you build a really big backlog of things that…”
“I've got one called tools, simonw/tools”
“the moment the coding agent has written the code, run the code, plotted a graph of how it worked or whatever, that's what turns it…”
From the episode
An AI state of the union: We’ve passed the inflection point, dark factories are coming, and automation timelines
Simon Willison