Workbase
2026 · open source · claude code + codex · Updated
Workbase is a notes folder for one job that my coding agent keeps up to date. I clone the company's repos inside it, and every agent session started below it reads the same manual, which says where each kind of fact goes and where to look before answering.
Figure 01 · Lookups and captures · workbase
Where the manual sends a question and where it files a capture
- Question
- Answer
- Capture
Where it came from
Before this I ran a heavier setup at my job for about four months: 27 skills, a phase pipeline per feature and an activity log. Most of it went unused. Twelve of the skills were never used in about 190 work sessions, and the hand-kept daily log and work board stopped being updated within a few weeks.
Workbase is what was left after keeping only the parts I used. It stores what the company's systems don't: why something was decided, what an acronym means, how the deploy broke last time and how it was fixed.
The folder and the manual
I make one workbase per job with bin/new-workbase and clone the company's repos inside it. Its .gitignore starts with /* and lists only the workbase's own files, so the repos are never committed with the notes. The script won't nest a workbase inside another git repo, overwrite an existing folder, or add a remote that matches a pattern I've marked as personal.
The manual is AGENTS.md, and CLAUDE.md is a symlink to it, so Claude Code and Codex read one file. Claude Code loads it in any session started in a repo below the workbase. An @AGENTS.md import doesn't work for this: Claude Code skipped the import from the parent folder when the session started inside a repo.
Where things go
The manual names a file for every kind of fact: a row in decisions.md with the date, what, why, who and a link; shipped outcomes with their numbers in wins.md; acronyms in glossary.md; a card per repo or service in systems/. When I mention one of these, the agent files it in the same session and says where in one line. Anything it can't place goes to inbox.md.
A second list says where to look before answering: errors and permission problems in systems/gotchas/ first, a feature's state in its project folder, an acronym in the glossary. Reads are free, but sending messages, editing shared docs and changing tickets wait for my yes.
Testing it on my notes
I wrote ten questions and capture requests from my own work notes, each with a known answer: seven lookups and three captures. Each ran on Sonnet against my old setup and against the same notes moved into the template, with every session started inside a repo. evals/grade.py marks them mechanically: a lookup passes if the answer contains the expected words, a capture if a file with the expected name changed and holds them.
The old setup passed 4 of 10. All three captures went into the agent's auto-memory, which nothing else can search. The list of shipped features existed only as JSON in a registry nothing indexed, and the doc index lived outside the workspace. The workbase passed 9, in 51 turns to the old setup's 65.
Both setups missed the same question: the fix for a known 403 was in a gotcha note, and neither session looked there. Adding "errors go to systems/gotchas/ first" to the manual fixed it. The question then passed in 3 of 3 reruns and took 4 turns instead of 12.
Each task ran once per setup, on one model, graded by keyword and file. The tenth pass came from a manual change made after seeing that miss, and it was checked on the same question. The tasks and notes are private, so the repo ships the scripts and an example task file for running the same test on your own notes.
What it doesn't do
It's built for one person: no shared registry, no team dashboard, and no notes committed into team repos. Nothing runs on a schedule.
The company owns the workbase, so it stays on company machines and company git, or on the laptop with --local, which keeps the remote but disables pushes. Not built yet: a morning summary from calendar, mentions and open PRs, and a review drafter that works from wins.md.
- manual
- 55 lines · a 10-row filing table
- setup flags
- --remote · --local · --adopt
- eval cost
- $0.98 old setup · $0.97 workbase
- migration
- dossiers 1,866 → 44 files · no lookup lost
- bundled skill
- council · runs only when asked
- runtime
- Bash setup script · Python grader
- status
- MIT · github.com/ong6/workbase
// AI TOOLING
All four layers
01 · Memory
Workbase
A notes folder per job that the agent keeps up to date.
02 · Context
Agent skills
The reusable instructions my agents load.
03 · Evaluation
Skillsmith
Drafts a skill, then a blind test decides whether it stays.
04 · Runtime checks
Groundplane
Checks agent output against recorded tool results.
// NEXT
Agent skills
The reusable instructions my agents load, kept in one public repo