Part butler,
part engineer

When you start using AI APIs seriously, the first bill can be a surprise. Tokens add up fast — especially if you are running the same capable model for every single task, including the trivial ones.

The key insight is that different tasks need different models. A quick nudge message asking “have you worked out today?” does not need the same horsepower as a complex reasoning task or match day analysis. Using Claude Haiku for background nudges and summaries, and reserving Claude Sonnet or Gemini Pro for the heavy lifting, cuts costs significantly without any loss in quality where it matters.

OpenClaw makes this easy with per-cron model overrides. You define the model at the job level, so routine tasks run cheaply and important ones get the full treatment.

Another big lever is context size. Woodhouse loads memory files at the start of each session. The more you dump into memory, the more tokens every conversation costs. Keeping MEMORY.md lean and purposeful is not just good housekeeping — it is cost management.

My current split: Haiku for heartbeat nudges and routine checks, Sonnet for conversations and project work, Gemini 2.5 Pro for match day analysis and deep research. The difference in monthly spend is meaningful.

A conversational AI that only responds when you talk to it is useful. An AI with routines that run whether you prompt it or not is something different — more like a colleague who is actually on top of things.

Woodhouse runs a set of scheduled jobs that cover the things I want monitored without having to remember to ask. Fitness nudges go out at 18:30 on weekdays. A security log review runs each morning. Middlesbrough FC news gets scraped at 08:00 and delivered at 12:00. On match days, a full analysis — form, stats, injuries, H2H — is ready by 11:00.

OpenClaw uses two mechanisms for this: heartbeats and cron jobs. Heartbeats are a recurring check-in where Woodhouse decides whether to say something. Cron jobs are precise — exact times, specific models, isolated sessions. For things where timing matters, cron is the right tool. For “check if anything needs attention” — heartbeat.

The design principle I settled on: routines should feel like a colleague who noticed something, not an alert system blasting notifications. The filter is “would I actually want to know this right now?” If not, stay silent.

Woodhouse lives in OpenClaw, but the project brain lives in Obsidian. The combination turned out to be more powerful than I expected.

Obsidian is a local Markdown note-taking app with a graph view and powerful linking between notes. Every project gets a folder. Documentation, decisions, article drafts, change logs — all in one place, nicely structured, and available on every device through sync.

OpenClaw reads from the workspace directory on the Shuttle PC. I set up a sync manifest so Obsidian files mirror across to the workspace — meaning Woodhouse can read project documentation directly from Obsidian without me having to paste things into the chat. I just say “read the Spearhead docs” and he finds them.

The combination also solves the memory problem. AI assistants start fresh each session. Obsidian holds the project knowledge that would be too large for the AI's context window — detailed architecture notes, full change logs, article drafts. The AI reads what it needs, when it needs it.

Managing the whole thing from Telegram means I can update documentation and brief Woodhouse on projects from my phone, wherever I am. That mobility turned out to matter more than I expected.

The single biggest factor in how useful Woodhouse is day-to-day is not the AI model — it is the instructions that define how he thinks and behaves. Get this right and even a smaller model performs surprisingly well. Get it wrong and a frontier model will still frustrate you.

Woodhouse has a SOUL.md file — literally a document describing his personality, values, and style. Sharp. Direct. Dry. No filler. Trusted colleague energy, not eager assistant. This sounds abstract but it has a real effect on every response he gives. When I changed “be helpful” to “be genuinely helpful, not performatively helpful”, the quality of responses noticeably improved.

Beyond personality, AGENTS.md covers operating rules: when to ask before acting, how to handle memory, what counts as a hard line. The hard rules section is particularly important — things like “never commit to GitHub without explicit permission in the current conversation” exist because I learned the hard way what happens without them.

The takeaway: spend time on your instruction files. They are the most leveraged investment you can make. A well-crafted SOUL.md outperforms model upgrades for day-to-day quality.