Theorist
February 28, 2026 ยท View on GitHub
Maintain a continuously updated operating theory for ongoing engineering work.
theorist is an agent skill (Codex and Claude compatible) that keeps a per-repo narrative document at
THEORY.MD (repo root). The document explains the problem thesis, the current
system model, why the strategy is shaped the way it is, and where uncertainty
still exists.
It is intentionally not a task plan, changelog, or status report.
Install
Use a skills directory for your client and clone into it:
SKILLS_DIR="${CODEX_SKILLS_DIR:-${CLAUDE_SKILLS_DIR:-$HOME/.codex/skills}}"
mkdir -p "$SKILLS_DIR"
git clone https://github.com/blader/theorist.git "$SKILLS_DIR/theorist"
If you already have the repo checked out, copy just the skill file:
mkdir -p "$SKILLS_DIR/theorist"
cp SKILL.md "$SKILLS_DIR/theorist/"
Usage
Invoke it using your client's skill command.
Or ask naturally:
- "Keep the operating theory updated while you work."
- "Rewrite the theory doc based on what we just learned."
- "What changed in our theory after this debugging pivot?"
What it enforces
- One theory doc per repo at
THEORY.MD(repo root) - Always-on activation in every session
- Sticky activation: once active, it stays active for the full session
- Holistic rewrites (not appended logs)
- Clear separation between:
- problem thesis,
- operating theory,
- strategy,
- key discoveries/pivots,
- open questions.
- Frequent in-session updates as learnings arrive (especially after investigate/implement/verify loops and verification outcomes).
- Updates when understanding changes, not when code churn happens.
Why this helps
Most agent workflows optimize for local task completion. theorist adds a
lightweight strategic memory layer so future decisions remain coherent with the
latest mental model of the system.
Files
SKILL.md: canonical skill behavior and writing standard
License
MIT