OPERATOR · Cheat Sheet

The full protocol is a system. Reading it linearly takes about 25 minutes. Living it takes weeks. This track orders what to install when, so you build discipline before you exhaust enthusiasm.

After Day 30: continue weekly review, monthly strategic, quarterly meta. Re-read Anti-Patterns any time you feel something is silently off.

Four surfaces. Each has a job. Mis-routing burns tokens and loses context.

02.1Session opener

For continuation work, use a standard trigger phrase (e.g. "I'm back" or "continuing {project}"). Required first action: read the project journal, then the latest session log entry, then orient before acting.

02.2Session closer

Append a journal entry. Required fields:

• Date
• What shipped (concrete deliverables, files touched)
• Decisions made (explicit, not implicit)
• Open threads (carry-forward items)
• Next action

This is the only mechanism that preserves context across sessions. Skip it and the next session pays the re-priming tax.

02.3Session length discipline

Long sessions degrade. Symptoms: drift to summarization, loss of specificity, repeated re-confirmation of established facts, hedging where there was none. When detected, close, journal, restart. Marginal value of session minute 180 is below minute 1 of a fresh session.

03.1Per-project structure

{project_root}/
  CLAUDE.md                  # project-level instructions, auto-loaded by Claude Code
  CLAUDE_JOURNAL.md          # append-only session log
  cowork/                    # Cowork project scope (if applicable)
  .claude/skills/            # project-scoped skills (Claude Code)
  docs/                      # human-readable documentation
  src/                       # code (if applicable)

03.2What belongs in CLAUDE.md

• Project glossary (canonical terms, banned drifted terms)
• Stack overview
• Conventions Claude must follow (file structure, naming, style)
• Known gotchas
• Pointer to journal

03.3What does NOT belong

• Session logs (use journal instead)
• Full strategy docs (link, do not embed)
• Anything that changes weekly (it will go stale)
• Personal preferences (use user preferences setting)

Skills are the only reliable mechanism for forcing canonical vocabulary, voice, or visual tokens across sessions. Prompt-level rules decay. Skill descriptions auto-fire on task triggers.

04.1When to write a skill

• Domain vocabulary Claude pattern-matches away from
• Brand voice conflicting with defaults
• Visual design tokens for recurring visual work
• Multi-step protocols repeatedly re-explained

04.2When NOT to

• One-off work
• Anything that fits in a single prompt and won't recur
• Personal communication style (use user preferences instead)

04.3Description discipline

The description field is the trigger. Vague descriptions mean silent miss. Test by giving Claude a task in a clean session; if the skill does not load, the description is broken.

05.1Decision compilation

Pass 1: Compile strongest case for {option}. Identify load-bearing assumptions.
Pass 2: Triangulate against {independent angle}. Identify contradictions.
Verdict: GO / PASS / PIVOT. Operative recommendation in one sentence.

Single-pass analysis on load-bearing decisions produces frame lock-in. Multi-pass is mandatory above a defined stakes threshold.

05.2Research synthesis

Source set: {explicit list, or "web search"}
Question: {one question, falsifiable}
Output: {claim, evidence, confidence} per finding
Defer any finding below {confidence threshold}

05.3Build session

Context: {pointer to journal / CLAUDE.md — do not re-explain}
Today's deliverable: {one concrete thing}
Acceptance criteria: {testable}
Out of scope: {explicit list}

The out-of-scope line is the highest-leverage element. Without it, agentic surfaces expand work into adjacent areas by default.

05.4Stress-test

Position: {claim}
Adversarial frame: {opposing camp / failure mode}
Output: strongest objection, weakest objection, evidence that would settle it

06.1Source hierarchy

• Current context — what is already in conversation. Free. Use first.
• Memory — cross-conversation. Free. Recency-biased. Partial.
• Project files — CLAUDE.md, journal, repo. Read, do not ask.
• Web search — anything time-sensitive or post-cutoff.
• Ask the user — last resort. Batch questions. Never one at a time.

06.2Search budget

Good queries: 2–5 words, content-noun-heavy, no operators unless required.

07.1Pre-flight

• Journal read
• Today's scope written down (one deliverable, not a list)
• Out-of-scope list written down
• Acceptance criteria defined

07.2During

• Agentic surfaces overrun scope by default. Reject expansion proactively, not retroactively.
• Stop and journal at logical breakpoints, not when tired. Tired journal = degraded journal = worthless.

07.3Post-flight

• Journal entry: files touched, decisions made, open threads
• Non-critical follow-ups → backlog file, not next session
• If something broke and wasn't fixed, journal it explicitly. Silent breakage compounds.

Research is unbounded. Build is bounded. Mixing them produces neither.

08.1Structure

• Question framed before starting (falsifiable)
• Sources logged as encountered
• Synthesis at end, not running commentary
• Verdict: actionable / not actionable / needs more evidence

Skipping the weekly review is the single most common discipline failure. It is the only cadence that catches scope drift before it compounds.

10.1What it is

Cross-conversation context inserted at runtime. Recency-biased. Summarized and lossy.

10.2What it is not

• A reliable archive
• A search index
• A substitute for project files

10.3When to edit

• A stored fact is wrong → correct it
• A sensitive topic should never resurface unprompted → exclude it
• A status changed (job, location, relationship) → update it

Do not store: secrets, credentials, anything that should not surface in an unrelated future conversation.

A protocol without measurement is aspirational. These are the metrics that indicate whether the protocol is being followed. Measure causes, not effects.

Twelve known failure modes. Each compounds silently if uncorrected.

Before treating Claude output as authoritative, verify:

If any answer is no, the output is provisional. Mark it.

• Context window — large but not infinite. Degrades before truncating.
• Memory — cross-conversation, recency-biased, not a substitute for journaling.
• Skills — must be loaded by description match. Vague descriptions = silent miss.
• Web search — top-10 snippet-level. Use web_fetch for full pages.
• Tool routing — Claude does not auto-pick the right surface. Operator does.
• Agentic surfaces — scope expansion is the default behavior. Constrain explicitly.

Run this audit on the cadence indicated. Tap to check items off — state is local-only and resets on reload (intentional; the audit is a ritual, not a record).