Start with the top of the chart if you are drowning in separate tools: pewdiepie-archdaemon/odysseus led the week with 1,862 star events as a self-hosted workspace that combines AI chat, research, documents, notes and planning. The experiment is simple: for three days, route everything you would normally scatter across four apps into it, then honestly check which features hold up. If the planning view is weak, keep using it only for chat plus documents and no more.
If your daily problem is an assistant that starts coding before it understands you, two entries near the middle pair well. mattpocock/skills contains small, editable workflows, including one that asks you detailed questions before any implementation. affaan-m/ECC offers a broader set of routines for planning, tests, review and remembering decisions. Take one skill from each, rewrite it to match your project’s terminology, and see whether your agent’s first attempt at a feature actually matches the requirement this time.
For people who work with piles of documents rather than code, microsoft/markitdown is the quietly practical entry: it converts PDFs and Office files into Markdown so a mixed folder becomes searchable text. Run it on your ten messiest files first, not clean ones, because that is where heading and table extraction actually breaks. RyanCodrai/turbovec is relevant once you have generated embeddings for the documents: it stores and queries those vectors, and lets you add more without a separate index-training pass.
The memory theme shows up again in NousResearch/hermes-agent, a personal agent that saves completed tasks as reusable skills and lets you search old conversations. That is a different promise from mattpocock/skills: one tries to make the agent remember, the other tries to make it ask. If your work is recurring, like weekly reports or repeated refactors, hermes-agent’s persistence is worth the trial; if it is mostly one-off features, the questioning skill is the smaller workflow change.