# Lemon Agent > A practical guide to working with coding agents: a short starting course, > three deeper field guides, two numbered failure lists (installable as > agent skills), and services agents can call directly. House rule: nothing ships without a working answer to "how would a user actually use this, today?" Everything is v0; guide step slugs and list entry ids (PF-NN / AF-NN) are stable and citable. Provenance: mostly AI-generated, human-guided and reviewed. ## Start here The human-first onboarding course is at /start/: seven short lessons on scoping, planning, guiding, verifying, and reviewing a first coding-agent task. Each lesson includes a decision exercise (same data format as the interactive examples below, raw JSON at /examples/.json) and a copyable action. - [Lesson 1: From answers to actions](/start/from-chat-to-agent.md) - [Lesson 2: Choose a safe first task](/start/choose-a-safe-task.md) - [Lesson 3: Write the brief](/start/write-the-brief.md) - [Lesson 4: Plan before code](/start/plan-before-code.md) - [Lesson 5: Watch the work](/start/watch-the-work.md) - [Lesson 6: Verify the result](/start/verify-the-result.md) - [Lesson 7: Review and save](/start/review-and-save.md) ## The guides Adopt one pattern at a time. Each chapter ends with an optional prompt you can adapt to the coding agent you use. Each guide's step 0 is its baseline in one paste. Guide index: /guides/ ### Working with Coding Agents (№ 3), raw at /guides/coding-agents/*.md - [Step 0: The one-paste baseline](/guides/coding-agents/full-setup.md) - [Step 1: Make each correction stick](/guides/coding-agents/working-agreements.md) - [Step 2: Prove it before and after](/guides/coding-agents/verification-loop.md) - [Step 3: Make the agent interview you](/guides/coding-agents/design-before-code.md) - [Step 4: Reviews that try to kill the work](/guides/coding-agents/adversarial-review.md) - [Step 5: Fan out, then funnel](/guides/coding-agents/fan-out.md) - [Step 6: Choose the right tool surface](/guides/coding-agents/tool-surfaces.md) - [Step 7: Build lots of small tools](/guides/coding-agents/build-tools.md) - [Step 8: Views on demand, not apps](/guides/coding-agents/jit-ui.md) - [Step 9: Turn checkable rules into checks](/guides/coding-agents/graduate-to-hooks.md) ### Shipping Software with Git + Coding Agents (№ 4), raw at /guides/git/*.md - [Step 0: The one-paste repo baseline](/guides/git/full-setup.md) - [Step 1: Checkpoint commits](/guides/git/commit-small.md) - [Step 2: Branches bound the blast radius](/guides/git/branch-blast-radius.md) - [Step 3: Write history for the next reader](/guides/git/history-for-the-next-reader.md) - [Step 4: Review the diff, not the summary](/guides/git/review-the-diff.md) - [Step 5: Recovery is cheap. Use it.](/guides/git/cheap-recovery.md) ### Keeping a Vault (№ 5), raw at /guides/vault/*.md - [Step 0: The one-paste vault](/guides/vault/full-setup.md) - [Step 1: Plain files in a repo](/guides/vault/plain-files.md) - [Step 2: Location is state](/guides/vault/location-is-state.md) - [Step 3: Capture now, sort later](/guides/vault/capture-now.md) - [Step 4: The daily log](/guides/vault/daily-log.md) - [Step 5: Teach the agent the shape](/guides/vault/teach-the-shape.md) - [Step 6: Run a doctor](/guides/vault/run-a-doctor.md) ## Interactive examples Branching mini-conversations: pick the agent's next move, see where it leads (baseline vs correct vs plausible-but-wrong, each with its downstream outcome and AF/PF citations). Every pattern chapter embeds one; each is also standalone at /examples// with raw JSON at /examples/.json: - coding-agents: clean-probe (verification) · unnamed-fork (design forks) · borrowed-rules (working agreements) · review-that-refutes (adversarial review) · vanishing-findings (fan-out) · cli-api-or-mcp (tool surfaces) · third-time-tool (small tools) · dashboard-that-rots (JIT views) · promise-vs-hook (hooks) - git: wall-of-diff (checkpoint commits) · experiment-on-main (branches) · weird-retry-loop (history) · merge-on-summary (diff review) · broken-main (recovery) - vault: app-or-files (plain files) · status-field-rot (location is state) · idea-mid-work (capture) · log-it (daily log) · twice-corrected (teach the shape) · doctor-first-run (doctor) - start (the course's lesson exercises, one per lesson): from-chat-to-agent · choose-a-safe-task · write-the-brief · plan-before-code · watch-the-work · verify-the-result · review-and-save ## Lists (the guides' watch-out reference) - [The Prose Failure List](/lists/prose-failure-list.md): 24 numbered failures of AI-assisted prose (PF-01…PF-24), run as a subtraction pass. Paste-ready prompt in the page's "How to use it" section. - [The Agent Workflow Failure List](/lists/agent-workflow-failure-list.md): 20 numbered failures of agent-assisted engineering (AF-01…AF-20), each grounded in measured work. Entries marked ⚙ can be checked partly by a script. ## Craft (incubating, unlisted) Domain applications of the same method, indexed at /craft/ (raw markdown at /craft/.md). Craft pages incubate outside primary navigation until they earn it (installable instrument + judgment artifact + worked path + external use). Statuses are honest: v0 = something on the page is usable today; scoping = the published plan, stated before the work. - [Lemon Agent for Writing](/craft/writing.md): the Interview Conductor (installable `writing-interview` skill; instrument page at /writing/interview-conductor/, raw at /craft/interview-conductor.md) plus the PF subtraction pass. v0. - [Lemon Agent for Knowledge Work](/craft/work.md): the verify-info pass producing a user-owned claim ledger; the installable skill is scoped on the page. v0. - [Lemon Agent for Audio](/craft/audio.md): audit a bounce against a reference with common tools (FFmpeg, librosa, pyloudnorm) and get a measurement-backed bounce-audit.md; the smpl toolkit is not yet installable and the page says so. v0. - [Lemon Agent for Science](/craft/science.md): a reproducible-project guide (Markdown + SQLite + Python; uv, sqlite-utils, Jupytext, Quarto) usable today; the pinned astronomy demonstration is in development. scoping. ## Install as skills Use `npx skills` for a managed install shared by Codex and Claude Code: `npx skills add chronick/lemon-agent --global --agent codex claude-code --yes` The repository contains: - `skills/prose-failure-list`: subtraction pass over a draft - `skills/agent-workflow-failure-list`: verification pass over a session, PR, or report - `skills/writing-interview`: the Interview Conductor, a one-question-at-a-time interview that ends in structured notes and a writing brief (never drafts for the user) Run the lists as advisory passes: they report, they never gate. ## Services - [pathgrip](https://pathgrip.net): a bookmark + provenance store with content-addressed snapshots, drift detection, per-agent bearer tokens, and an MCP surface. Operator instance today; public docs and self-serve keys are the next ship. ## Machine catalog - [catalog.json](/catalog.json): every member with tier, kind, status, chapters, install path, and URLs.