cabaynes/charles-agent-skills

Seven skills that I've built that are fundamental to my agentic workflow. Built for both **Claude Code** and **OpenAI Codex**. Both tools use the same open Agent Skills format, so every skill here ships in two versions: one per tool, built from a single source so they never drift apart.
- **Claude Code:** evaluated against Anthropic's writing-skills + skill-creator rubric, with **100% trigger accuracy** on a 20-query precision/recall benchmark per skill.
- **Codex:** every skill tested end to end through the real Codex CLI (**42/42** checks), plus a spec check and a behavior rubric.
> Renamed from charles-claude-skills on 2026-09-29, when the Codex versions shipped. Old links redirect here. The Claude skills moved from skills/ to claude/.
Pick your tool
| You use… | Install from | Run a skill with | Project instructions file |
|---|---|---|---|
| **Claude Code** | claude/ → ~/.claude/skills/ | /putdown | CLAUDE.md |
| **OpenAI Codex** | codex/ → ~/.agents/skills/ | $putdown | AGENTS.md |
| **Both** | Install each folder into its own tool | either | Handoffs in a repo's .putdowns/ work across both tools |
**How the repo is laid out:**
**If you're an agent installing these for someone:** find out which tool they use, copy the matching folder's skill directories into that tool's skills folder (table above), and tell them to start a fresh session. Don't mix them: the claude/ versions call Claude-only tools, and the codex/ versions call Codex's.
Skills in this repo
Names below use Claude Code's /name; in Codex, type $name.
**The Session Continuity Loop** — one folder, one set: openthreads, takenotes, putdown and pickup go together, in the order you meet them in a session. Docs: claude/session-continuity/README.md
Without this four-skill set, an agent's working memory resets with every fresh session. These four fill the memory gap between sessions and keep that memory healthy and complete, preventing duplicated mistakes and stale knowledge. Use openthreads and takenotes during a work session as you see fit, then end every session with a putdown. Start your next session with a pickup, and never repeat yourself again.
**/openthreads**
- **What it does:** Keeps the project's living to-do list in one file: parked ideas, decisions waiting on you, work nobody has picked up yet. Finished items become a 60-day log of outcomes.
- **Why I built it:** A handoff holds one session's next steps and memory holds facts. Neither held the backlog, so parked ideas lived in people's heads until someone re-read two weeks of notes.
- **When to use it:** As you work, whenever something gets parked: an idea for later, a question for someone else, a task no session owns yet. Putdown sweeps up anything you forgot.
**/takenotes**
- **What it does:** Saves what a session learned (decisions, dead ends, how the tools really behave) to permanent memory, then goes back over what's already stored and fixes anything that has become false.
- **Why I built it:** Every assistant can save new facts; almost none check whether old ones are still true. Stale memory is worse than none: it argues the agent out of work that's already done.
- **When to use it:** Any time you learn something that should outlive the session, not only at the end. A Putdown is a sticky note for tomorrow; a Takenote is a fact for forever.
**/putdown**
- **What it does:** Ends a session with a handoff note: where things stand, what's half-done, what must not slip, what was tried and failed. Then it saves everything to GitHub.
- **Why I built it:** A fresh session wipes the agent's working memory. Without a handoff, every session starts from zero and repeats the same dead ends.
- **When to use it:** At the end of every working session, or when the context window is filling up. It runs Takenotes first and Openthreads last, so one command saves all three.