notessamuele onelia

Claude Code

Working with AI coding agents in practice: workflows, configuration artifacts (skills, subagents, CLAUDE.md), and the human costs and limits of orchestrating them. Centered on Claude Code but the patterns generalize to any agentic coding tool.

Workflow Principles

Give the Agent a Way to Verify Its Own Work — Boris Cherny’s most-repeated advice: a verification loop is worth a 2-3x quality bump Three Modes of Working With AI — vibe coding, pairing, by hand: match the mode to the task Delegate vs Collaborate as the Meta-Skill of Knowledge Work — Katie Parrott’s sibling taxonomy: sort tasks by autonomy level before opening the agent The Orchestration Tax - Cognitive Bandwidth Doesn’t Parallelize — agents scale horizontally, your judgment doesn’t; the hidden cost of parallelism Everyone Says Loop, Nobody Defines It — AI-coding vocabulary spreads faster than its meaning; copy the workflow, not the slogan Loops Make Sense Only With a Fixed Feedback Signal — the boundary for autonomous loops: a confined process with an objective score to optimize against A Plan Document Can’t Hold Your Product Vision — why fire-and-forget loops fail for real products: specs can’t carry judgment that isn’t formed yet Loops Are Priced for Unlimited Token Budgets — loop advice inherits the advocate’s cost structure; check the price tag before copying the workflow The Thousand Line Review Ceiling — automated review stops being a reliable signal past ~1k lines; small batches keep machine review honest Agent Loops Only Work When the Output Is Binary — atomic essay distilling the loop criterion: no machine-checkable score, no loop

Configuration Artifacts

Skills Are the Unit of Reusable Expertise — repeated prompts become versionable, shareable skills; institutional knowledge in git Subagents Keep the Main Context Clean — isolated contexts and scoped permissions; separate concerns into separate workers A Claude Code Subagent Is a Markdown File — the demystification: a working subagent is a 15-minute text file Compound Engineering - Only Document Mistakes — the CLAUDE.md discipline: every real failure becomes a permanent rule The Three Prompt Rule for Promoting Prompts to Persistent Context — the trigger for writing one: stop typing the third repetition and put it in a goal, skill, or CLAUDE.md Coding Agents Are Knowledge Work Agents in Disguise — why these primitives matter beyond engineering: a non-developer needs goals, skills, and CLAUDE.md just as much

In Practice

259 PRs in 30 Days, 2 Bugs - The Vanilla Setup — what disciplined agent use produces at full throughput Morning Agents from the Phone - Boris’s Full Daily AI Workflow — a full day built around parallel agent sessions Daisy’s 20-Claude Swarm Weekend — a 20-agent swarm building plugins over a weekend

See Also

  • @Boris Cherny — creator of Claude Code; his index holds the origin stories behind these patterns
  • @Agent-Native Software — the building-apps-for-agents counterpart to this using-agents cluster

Backlinks