> Source: https://www.airtightdesign.com/bob-builder/day-one/ # Day One Walkthrough **What your first session actually looks like, step by step. Read this before you start.** --- ## Before You Start You should already have: - [ ] All workspace files generated (CLAUDE.md, SOUL.md, USER.md, AGENTS.md, SCHEMAS.md, MEMORY.md, TOOLS.md) - [ ] Sequence counter and session log initialized (_sequence.md, _session-log.md) - [ ] Directory structure created (captures/, projects/, people/, memory/) - [ ] Empty indexes in place (captures/_index.md) - [ ] OpenClaw running with an agent runtime (Claude Code, Codex CLI, or similar) pointed at the workspace directory If you haven't done this yet, go back to `03-assembler.md` and complete the Assembly phase. --- ## The First Conversation Open your AI tool in the workspace directory. Your first message should be simple: ### You say: > "Wake up. Read your boot files and tell me what you see." ### What should happen: Your assistant reads the files in order (CLAUDE.md → SOUL.md → USER.md → MEMORY.md → AGENTS.md). It should: 1. **Acknowledge who it is** — by name (the name you chose) 2. **Acknowledge who you are** — your name and role 3. **Note that memory is empty** — this is a fresh start, nothing to remember yet 4. **Note that there are no captures yet** — empty indexes, no projects, no person files 5. **Write a session-start entry** to _session-log.md If the assistant does something generic like "Hello! How can I help you today?" — it didn't read the boot files. Tell it: "Read CLAUDE.md first, then follow the instructions there." ### What it should sound like: The tone should match what you configured in SOUL.md. If you set up a direct, no-nonsense personality, it shouldn't be bubbly. If you set up a warm, supportive personality, it shouldn't be cold. If the tone is wrong, correct it immediately: "You're being too formal. Read your SOUL.md personality section again — I want you [casual/direct/warm/whatever you configured]." --- ## Feed It Something Real Don't start with hypotheticals. Give it something from your actual day. ### Option A: Give it emails to triage > "Here's what's in my inbox this morning. Triage these for me." Then paste 5-10 emails (or if you have Gmail MCP configured, tell it to read your inbox). Watch what it does: - Does it classify correctly? (Questions, tasks, commitments, context) - Does it create person files for people it encounters? - Does it update the index? - Does it recognize what's urgent vs. what can wait? **Expect mistakes.** This is the first run. The assistant has rules but no pattern memory yet. Correct every mistake: - "That's not a task, it's a commitment — they promised us something, we didn't promise them." - "Don't create a capture for that — it's a newsletter, just archive it." - "Good catch on that commitment. But the due date is Friday, not next week." ### Option B: Describe your current situation > "Here's what's on my plate right now. I have three active projects: [name them]. My biggest headache is [describe it]. The people I work with most are [name them]." Let the assistant ask follow-up questions. Watch it: - Create project directories and _project.md files - Create person files for the people you mention - Capture context about your current situation - Start building indexes ### Option C: Walk through your calendar > "Here's my schedule for today. Walk me through what I should be thinking about." If Calendar MCP is configured, tell it to read your calendar. Otherwise, paste your schedule. Watch it: - Identify prep work needed for meetings - Flag commitments due today - Note who you're meeting with (should cross-reference or create person files) - Suggest follow-ups from yesterday --- ## The Correction Loop This is the most important part of Day One. **You are training your assistant.** Every correction teaches it something. ### How to correct classification: **Bad:** "That's wrong." **Good:** "That's not a task — it's a commitment. When Jake says 'I'll get the spec to you by Thursday,' he's making a promise. File it as a commitment with Jake as 'Promised by' and me as 'Owed to.'" The "good" version teaches a pattern, not just a correction. ### How to correct tone: **Bad:** "Be different." **Good:** "You're being too verbose. I asked a simple question — give me the answer in one sentence, not three paragraphs. Save the detail for when I ask for it." ### How to correct autonomy: **Bad:** "Don't do that." **Good:** "Don't create a project without asking me first. Captures are fine to create silently, but projects need my approval — I want to name them and scope them myself." ### How to correct information: **Bad:** "That's not right." **Good:** "Jake isn't a client, he's on my team. Move his person file from people/clients/ to people/employees/. And update his title — he's a Senior Engineer, not a contractor." ### After each correction: Ask: "Update your MEMORY.md with what you just learned, so you don't make the same mistake next session." Not everything needs to go in memory — only patterns that should persist. "Jake is an employee, not a client" is a correction to the person file, not a memory entry. "When I say someone's name without a company, assume they're on my team" is a pattern worth remembering. --- ## What to Do in the First Hour Here's a realistic first session agenda: ### Minutes 0-10: Boot and verify - Assistant reads boot files - Verify it knows its name, your name, and its framework name - Verify the tone feels right - Fix anything that's off ### Minutes 10-30: Feed it real data - Give it emails, your calendar, or a description of your current work - Let it create captures, person files, and (maybe) projects - Correct classification mistakes as they happen - Don't worry about perfection — volume matters more than accuracy right now ### Minutes 30-45: Review what it built - Ask: "Show me the captures index." - Ask: "What person files did you create?" - Ask: "What's in MEMORY.md now?" - Correct anything that's wrong. Reclassify, move, merge, or delete. ### Minutes 45-60: First briefing test - Ask: "What needs my attention right now?" - Evaluate the briefing structure — is it what you expected? - If sections are missing or wrong, tell it: "I want the briefing structured as [your preference]. Update AGENTS.md reporting section." - Ask: "What open questions do I have?" and "What commitments are due this week?" ### Before you close: - Ask: "Write your session-end entry to the session log." - Verify it lists all the captures it created and files it modified. - If it misses something, tell it. --- ## What "Good" Looks Like After Day One By the end of your first session, you should have: - [ ] **5-15 captures** — a mix of types, mostly from whatever you fed it - [ ] **3-10 person files** — for the people who appeared in your captures - [ ] **0-3 projects** — only if the work was clearly project-shaped - [ ] **An index that reflects reality** — every capture is listed, statuses are correct - [ ] **A session log with one session-start and one session-end entry** - [ ] **A MEMORY.md with 2-5 entries** — things it learned from your corrections - [ ] **An assistant that sounds like what you configured** — right tone, right verbosity What "good" does NOT look like: - 50 captures (too aggressive — it's capturing noise) - 0 captures (too conservative — it's not catching anything) - Everything is a task (classification isn't working) - The assistant sounds generic (SOUL.md isn't being applied) - No session log entry (continuity is broken) --- ## Common Day One Problems **"It's not reading the boot files"** → Make sure CLAUDE.md is in the workspace root and your AI tool is pointed at that directory. Some tools need to be told explicitly: "Read CLAUDE.md and follow its instructions." **"It's being way too verbose"** → Correct immediately: "Too many words. Be concise. If I need more detail, I'll ask." Then update SOUL.md if the verbosity setting isn't strong enough. **"It created 30 captures from 5 emails"** → It's being too aggressive. Tell it: "You're over-capturing. A single email usually produces 1-2 captures, not 6. Only capture what's actionable or worth remembering." **"It didn't capture anything"** → It's being too conservative. Tell it: "You missed a commitment in that email — [person] said [thing]. You should have caught that. Default to capturing, and I'll tell you when it's too much." **"It keeps asking me what to do"** → Tell it: "For reads, captures, and file organization — just do it and tell me what you did. Only ask me before sending messages or taking external actions." **"The directory structure is wrong"** → Fix it now before captures accumulate: "Move [directory] to [correct location]. Update any references." --- ## Day Two and Beyond Day Two is where continuity gets tested. When you start a new session: 1. The assistant should read its boot files automatically 2. It should read the session log and know what happened yesterday 3. It should pick up where you left off — same captures, same projects, same people **If it doesn't remember yesterday,** check: - Did it write a session-end log entry? (If not, the new session has nothing to absorb) - Are the capture files actually saved? (Check the directories) - Is it reading _session-log.md at startup? (Check the boot sequence in CLAUDE.md) **Day Two priorities:** - Feed it more data (email, messages, meetings) - Let it build on yesterday's captures (update existing ones, not just create new ones) - Test a briefing: "Give me my morning briefing" - Continue correcting — Day Two should have fewer mistakes than Day One By **Day Five,** the assistant should be classifying correctly most of the time, your key people should have person files, and the briefing should be genuinely useful. If it's not, re-read your AGENTS.md and SOUL.md — the configuration might need tuning. By **Week Two,** you should trust the assistant enough to stop checking every capture. The corrections become occasional rather than constant. By **Week Three,** the assistant knows your world. It catches things you didn't. That's when it starts being indispensable. --- _The first day is the hardest. It gets dramatically better from here. Be patient, be specific with corrections, and trust the process._