Small, portable Skills for Claude Code, Pi, Codex and OpenCode. Work with evidence, keep state only when needed, and avoid unnecessary agent overhead.
Bet small before you spend big. One safe, tiny, falsifiable bet can save a costly deep dive—but only if the possible upside warrants it. Otherwise go NORMAL; if already correct, ABSTAIN. Evidence: across two local Muse studies (126 runs), v0.4 had lower mean wall time than the tested controls. In the 72-run ablation, paired time fell 17.9% vs Efficiency and 18.0% vs No-BET, comparing both-PASS pairs. Read results and limitations →
npx skills add tc3oliver/skills --skill betfirstAgent-guided install → · Explore BetFirst → · 繁體中文 → · How it works →
| Skill | What it helps with |
|---|---|
| BetFirst | Make Cheap Bets. Win Big. Lose Small. Tiny feasibility probes before expensive research and PoCs. |
| Backlog | Persistent coding tasks, ready queue, and handoff evidence—without Backlog.md. EN / 繁體中文. |
| audit-agent-instructions | Find duplication, instruction drift and excessive context. |
| diagnosing-bugs | Reproducible debugging and focused regression tests. |
| writing-for-agents | Clearer, more efficient instructions for coding agents. |
| skill-optimizer | Skill routing, prompt size and reliability. |
| resolving-merge-conflicts | Resolve conflicting code without losing either side's intent. |
Agent Skills CLI (choose the agent and installation scope interactively):
npx skills add tc3oliver/skillsClaude Code plugin:
/plugin marketplace add tc3oliver/skills
/plugin install tc3oliver-skills
Plugin entries use /tc3oliver-skills:betfirst, /tc3oliver-skills:backlog, etc. Portable Skill installations expose each SKILL.md through the compatible agent's discovery system; avoid installing the same Skill twice.
BetFirst decides whether a small probe is worth the risk. Backlog owns persistent tasks and evidence. Claude Team Kit manages explicit native Agent Teams. They work independently or together without a new orchestrator. Composition guide →
Backlog is a lightweight persistence CLI, not a background orchestration service:
/backlog add "Build OAuth login"
/backlog next
/backlog do BL-7E18C75F3D31A0B2
/backlog status
Backlog CLI reference → · Legacy v1.6 archive →
Earlier BetFirst revisions were slower on some simple tasks. v0.4 was changed to NORMAL by default. Published numbers are local synthetic-task observations, not a universal or causal guarantee. Benchmark report → · Experiment history →
python3 -m unittest discover -s research/backlog/tests -v
python3 -m unittest discover -s research/betfirst/tests -vMIT License. Existing upstream material and adapted sources retain their original notices.