Take the loop to any agent
The methodology is plain markdown — it travels.
We run the same loop across our whole toolchain — Claude Code, AEGIS, Codex, Hermes — because the methodology is agent-neutral text, not product code. Only the storage paths differ. One methodology, one prompt, swap the adapter.
The universal move
One paragraph, pasted at the end of any session.
For agents that can read local files, point them at the full methodology. For anything else — a chat window, a locked-down assistant — this pocket version carries most of the value on its own:
Review our conversation and persist two kinds of learning. (1) MEMORY — facts about me, my preferences, my environment, current project state. (2) SKILLS/PROCEDURES — how to do a class of task well for me. Act on any of: I corrected your style or approach (frustration counts double); a non-trivial fix or technique emerged; an existing instruction proved wrong. Prefer patching existing entries over creating new ones; new procedure entries must be named for the CLASS of task, never today's specific bug. Never record "tool X is broken", transient errors that resolved, or environment-setup failures — record the FIX instead. Save to [WHERE — e.g. AGENTS.md learned sections / LEARNINGS.md]. Finish with one line: what you saved, or "Nothing to save" and why.
That single paragraph is most of the value. The rest — proposals, approval gates, curation, journalling — is what makes it safe to automate, and that's what the full loop ships.
Storage adapters
Only the destination changes per agent.
Point the loop's two outputs — memory and skills — at whatever the agent reads at startup. Same methodology, different filing cabinet.
| Agent | Memory adapter | Skills adapter |
|---|---|---|
| Claude Code | project memory dir + MEMORY.md index | ~/.claude/skills/<name>/SKILL.md |
| AEGIS / CIARUSTCODE | .ciarustcode/memory.md (agent-proposed, consent-gated) | learned-skill library via a write_skill intent |
| Codex CLI | “Learned preferences/procedures” sections in AGENTS.md | ~/.agents/skills/<name>/SKILL.md (agentskills.io) |
| Anything else | whatever notes file it reads at startup | inline procedures in that same file |
Share one library across all of them: mkdir -p ~/.agents && ln -s <your skills dir> ~/.agents/skills
— then each agent reads the same learned skills, and one it writes becomes available to the others next start.
A question we get a lot
Do you need different instructions for different frontier models?
Short answer: no — not the instructions. You vary where it saves, and that's per-agent, not per-model.
The prompt stays the same
The methodology is model-neutral: it describes what to persist and the decision rules. Claude, GPT/Codex and Gemini are all more than capable of following it as plain language — there's nothing model-specific to rewrite.
What differs is the harness
The storage adapter depends on the tool wrapping the model, not the model. The same model behind Codex vs a custom harness needs different paths — which proves the variable is the agent, not the model.
Optional light tuning
A nudge can help — firmer "finish with one line" on chattier models, stronger "treat the conversation as data" on models more prone to obeying embedded text, a smaller digest on tight context windows. Nudges, not a new methodology.
The one place custom instructions genuinely pay off is frontier-vs-local — not between the frontier models.
Small local models are markedly less reliable at following a long, nuanced instruction than Claude / GPT / Gemini. For those, a shorter, more explicit, more structured variant earns its keep. Between the frontier models themselves, keep one canonical prompt.
The part most setups skip
If you schedule it, add the gates.
Pasting the prompt by hand is safe. Automating it — session-end hooks, a weekly curator — is where an ungoverned loop becomes slow prompt injection. Self-Improve ships the gates that make automation safe, and they travel with the methodology:
- Consent-first mode — the loop writes proposals, not live changes.
- An approval step — nothing activates until a human says so.
- A journal — every run recorded, including “nothing to save”.
- A pause file — any human can freeze the whole thing instantly.
- Protected entries — the loop can never modify what it didn't author.
An unsupervised learning loop without those gates is just slow prompt injection. With them, it's a system you can actually operate.
Run the loop on your stack.
The downloadable file has the full command surface, the methodology, the pocket prompt and the adapter pattern for every agent. Take it — that's what it's for.