PART 5 SYLAS LYU · FM-01 · REV B

APPLYING IT

Workflow recipes, safety, troubleshooting, a four-week plan and a cheat sheet.

SECTION 16

Workflow playbook

Pick the tool by where the inputs live and what has to happen to them. Reading and writing work belongs in claude.ai Projects with skills; Claude Code takes over once code or data files on your computer are involved.

The decision rules

  1. Inputs you can upload, output is text, a doc or a deck: claude.ai, inside the right Project.
  2. Inputs are files on your computer and something has to run on them: Claude Code in that folder.
  3. You keep copying information out of another app: connect it (connector in claude.ai, MCP server in Claude Code).
  4. You have explained the same procedure three times: write a skill.
  5. A side job would flood the conversation: a subagent.
  6. A rule must hold every single time: a hook or a deny rule, plus read-only files where a script could get round them; never only an instruction.
  7. The same job recurs on a timetable: a scheduled task.

RECIPE 1 A Project per subject

  • One Project per ongoing subject, holding its reference documents, with the instructions from Exercise 7.
  • Skills you use there each get a distinct trigger phrase in their description, so the right one fires.
  • For a summary you will reread, ask for it as a doc; for a one-off check, keep it in chat.
  • Use incognito chat for anything you do not want feeding into memory.

RECIPE 2 Meetings to notes

  • Single meeting: upload the transcript to the Project and say "make notes from this meeting"; the meeting-notes skill (section 8) shapes the output.
  • A month's worth at once: Exercise 25's batch script in Claude Code.

RECIPE 3 Reading and writing a paper

  • A Project for the paper: upload the sources, and put your citation style and spelling rules in its instructions.
  • Ask for claims with page references, then spot-check a few against the PDFs. Retrieval can miss passages and Claude can misattribute.
  • Draft in a doc so you can edit in place and comment.
  • Before sharing it, run the citation-checker subagent (section 12) in Claude Code over the draft and a sources/ folder.

RECIPE 4 Data analysis

  • Work in Claude Code in the analysis folder, never by pasting data into chat.
  • Set up once: a CLAUDE.md like the section 10 example; a deny rule on Edit(./data/raw/**) and read-only raw files (section 10); Python through uv.
  • For each analysis: start in plan mode, approve the plan, let Claude write the script, run it on a sample first, then the full data.
  • Read the code it wrote as well as the numbers. Ask it to print row counts and input file names so silent filtering shows up.
  • Before any data leaves your computer: if the dataset came with a data use agreement or privacy rules, check what they allow. Claude Code sends the contents of files it reads to Anthropic's servers to process them.

RECIPE 5 UX design work

  • Mockups and flows: ask for a design in claude.ai, which opens as an editable artifact; or work through the Figma and Canva connectors when the file lives there.
  • Critique: give Claude a screenshot or a Figma link and ask what to fix first and why.
  • Research synthesis: upload interview notes or survey results and ask for themes, each with the quotes behind it.

RECIPE 6 A website

  • Claude Code in the site's folder, with Git initialised so every change can be undone (git init once, then ask Claude to commit after each working change; Git basics are in section 4).
  • Plan mode for anything touching more than one page.
  • Ask Claude to open or preview the page after changes, then look at it yourself.

RECIPE 7 Long tasks that finish

This recipe adapts a setup posted by @beamnxw on X on 6 October 2026 ("Make Opus 5.5 Finish Long Tasks"), checked against Anthropic's docs on 7 October; corrections to the post are listed at the end. The example is a sourced write-up, such as a literature summary, in a folder holding task.md (the brief), sources/, drafts/ and final/. The aim is a run that leaves files you can open, evidence you can check, and a note the next session can resume from.

  1. Durable facts in CLAUDE.md, the task itself elsewhere:
MARKDOWN · C-65
# Write-up project
- Reference material is in sources/; working files in drafts/; approved files in final/.
- Plain prose, short paragraphs.
- Cite primary sources only, with the URL or page and the date checked.
- After each stage, update progress.md with decisions, open issues and the next action.
- When finishing, report the output paths and what was verified.
  1. The repeated procedure as a skill, in .claude/skills/write-draft/SKILL.md, run as /write-draft <topic>:
MARKDOWN · C-66
---
name: write-draft
description: Drafts a sourced write-up from the files in sources/ and checks every claim. Use when asked to draft or rewrite a write-up.
---
Topic: $ARGUMENTS

1. Read the relevant files in sources/.
2. Write an outline, then save the draft as drafts/draft.md.
3. Ask the evidence-reviewer subagent to check every factual claim against sources/.
4. Fix incorrect claims and mark unresolved ones with [CHECK].
5. Save the reviewer's table as drafts/checks.md.
6. Update progress.md, then report both paths and the verification results.
  1. A deny rule so finished work stays put (merge into .claude/settings.json). As section 10 explains, this stops Claude's own file tools; mark finished files read-only too if a script might touch them.
JSON · C-67
{
  "permissions": {
    "deny": ["Read(.env)", "Read(.env.*)", "Edit(final/**)"]
  }
}

Copy approved files into final/ yourself.

  1. A reviewer that can read and search but not edit, in .claude/agents/evidence-reviewer.md. The effort line gives it more reasoning than the main session (unless CLAUDE_CODE_EFFORT_LEVEL is set, which wins):
MARKDOWN · C-68
---
name: evidence-reviewer
description: Checks factual claims in a draft against the cited sources. Use after drafting.
tools: Read, Grep, Glob, WebSearch, WebFetch
effort: high
---
Read the draft and the sources it cites. For each factual claim, return one table row:
claim, verdict (verified, incorrect, unresolved), source location, correction needed.
For unresolved claims, say what evidence is missing. Never edit files.
  1. Start the session from the project folder and check the setup before the real run:
POWERSHELL · C-69
cd <project folder>
claude --effort medium

Inside it, run /context to see what loaded, /permissions to see the effective rules, and type @ to check that evidence-reviewer appears in the list. Medium is already the default effort for Opus 5.5. /effort high changes it mid-session and saves it as your default; press s in the /effort slider to apply a level to this session only.

  1. State what "done" means with /goal. After each turn a separate small model checks the condition against what the conversation shows, so ask for evidence in the conversation:
TEXT · C-70
/goal Use /write-draft on the topic in task.md. Done when drafts/draft.md and drafts/checks.md exist, every incorrect claim is fixed, unresolved claims are marked [CHECK], and both paths plus the verification results are shown in the conversation. Stop after 12 turns and report the blocker if this is not met.

The turn limit is judged by that checker from the conversation, so treat it as a soft limit; /goal clear (or /clear) removes a goal.

  1. Keep a handoff note, progress.md, updated after each stage:
TEXT · C-71
Task: <topic>
Outputs: drafts/draft.md, drafts/checks.md
Done: <finished stages and checks>
Decisions: <choices made, with their sources>
Open: <missing evidence or blockers>
Next: <one concrete next step>

To resume in a fresh session, ask: "Read progress.md, open the files it lists, continue from Next, and update the note."

Corrections and limits found when checking the post against the docs:

  • /agents in current versions only prints a reminder; the @-list or asking Claude shows which subagents exist.
  • Root CLAUDE.md is re-read after /compact, but path-scoped rules and nested CLAUDE.md files are summarised away until Claude reads a matching file again.
  • A top-level effortLevel in user settings does not apply to Opus 5.5; the CLAUDE_CODE_EFFORT_LEVEL variable, --effort, /effort and per-model settings do.
  • /rewind restores Claude's own file edits only; changes made by shell commands and by most subagents need git. Commit after each stage.
  • The post's "60% fewer tokens" comes from one customer quoted in the Opus 5.5 announcement; Anthropic's own figure there is about 40% lower cost than Opus 5 on typical workloads at default settings. Measure your own runs with /usage.
  • /goal needs a recent version of Claude Code (the docs give no minimum); run claude update, or switch from npm (section 9), if it is missing. It needs a trusted folder and is turned off by "disableAllHooks": true. A goal keeps your permission mode, so in Manual mode each turn still stops for approvals; press Shift + Tab to auto mode for an unattended run.

General habits

SECTION 17

Safety, privacy, cost and context hygiene

An agent acts with your permissions, so the main risks are things it does that you did not intend, data going somewhere it should not, and usage spent on context nobody needed. Each has a specific control.

Risks and controls

Risk Control
Claude edits or deletes the wrong files Manual or plan mode while learning; Git in every project; /rewind to roll back; deny rules and read-only files on folders that must not change
A command does damage Read prompts before approving; a PreToolUse hook for patterns you never want (section 13). Anthropic warns auto mode reduces prompts but does not guarantee safety; in auto mode broad allow rules such as Bash(*) are ignored, it pauses after 3 blocked actions in a row or 20 in total, and limits you state ("don't push") count as boundaries until compaction drops the message that stated them; use a deny rule for anything that must hold
bypassPermissions runs anything Only inside a disposable container or virtual machine, per Anthropic's permission-modes page
Hidden instructions in a web page, file or tool result (prompt injection) Connect only trusted servers; turn off write tools for Research; treat surprising actions as a signal to stop and check
A plugin, hook or MCP server misbehaves They run as you. Install from the official marketplace or Anthropic's directory; read hooks and "Will install" lists first
Secrets leak Never paste passwords or API keys into chat; deny reads of .env files (section 10); use narrowly scoped tokens for MCP servers
Data leaves where it is allowed to be Check data use agreements and privacy rules before any file reaches an AI service
A chat feeds memory you did not want Incognito chat; view and edit memory in Settings

For Anthropic's data handling terms, read the Claude Code "Data usage" page and the claude.ai privacy settings rather than relying on summaries, including this one.

What fills the context window

From Anthropic's "Extend Claude Code" page:

Feature Loads Cost
CLAUDE.md Whole file, every session Every request
Skills Name and description every request; full text when used Low until used
MCP servers Tool names at start; full definitions when needed Low until used
Subagents Their own separate context Isolated from your session
Hooks Nothing, unless they return output Zero

The same page warns that too much loaded context also makes Claude worse: skills trigger wrongly and conventions get lost. Disable add-ons you do not use.

Keeping usage down

Checking the output

Check these before you submit or rely on the output:

SECTION 18

Troubleshooting

Most failures come from five causes: a program not on PATH, invalid JSON, a file in the wrong folder, a session started before a change, or something switched off for that chat. Fixes below come from the Anthropic and tool documentation cited in section 21 unless marked otherwise.

Diagnostic commands

Command Shows
claude doctor (terminal) Install health, settings-file errors, update status
/doctor (in a session) Setup checkup that offers fixes
/status Version, account, model, which settings files loaded
/context What is filling context, including which CLAUDE.md files loaded
/mcp Each MCP server's status and tool count
/hooks Every configured hook and where it came from
/plugin, Errors tab Plugins that failed to load, and why
claude --debug-file <path> Full log of hooks, MCP and skill loading

Symptoms and fixes

Symptom Likely cause Fix
'claude' is not recognized Install folder not on PATH PATH lines in section 9, then a new window
'irm' is not recognized You are in CMD Open PowerShell
running scripts is disabled on this system (npx) PowerShell script policy Set-ExecutionPolicy -Scope CurrentUser RemoteSigned (Microsoft's about_Execution_Policies page)
Claude Code cannot find Git Bash Unusual Git install path Set CLAUDE_CODE_GIT_BASH_PATH in the env block of settings, as in section 5's example
"Settings Error" dialog at startup Invalid JSON in a settings file jq . <file>; look for trailing commas and comments
A setting seems ignored A higher-precedence file or flag sets it /status, then claude doctor
Skill never triggers Vague description, or YAML that fails to parse Add the words requests use; invoke with /name; check that /skills lists it and that --- is the file's first line
claude.ai rejects a skill upload: "Unexpected key(s)" Claude Code-only frontmatter field Keep to name, description, license, compatibility, metadata, allowed-tools
Uploaded skill never used in chat Toggle off, or code execution off Customize > Skills toggle; Settings > Capabilities
New subagent not found agents folder created after the session started Restart Claude Code
MCP server "Failed to connect" Server command fails on its own claude mcp get <name> for the Issue line; run the command by hand; $env:MCP_TIMEOUT = "60000" for slow first downloads
MCP server "Pending approval" Project .mcp.json server not yet approved Run claude interactively in that folder and approve it
MCP output cut short with a warning Over the 25,000-token default Raise MAX_MCP_OUTPUT_TOKENS, or ask for less output
MCP Inspector will not start Node older than 22.19.0 winget upgrade --id OpenJS.NodeJS.LTS (ID tested) or install a newer Node
Connector unavailable in a chat Toggled off for that chat, or signed out "+" > Connectors in the chat; reconnect in Customize > Connectors
Custom connector times out Server not reachable from the public internet Host it publicly, or use a local MCP server instead
Desktop app ignores a config change App still running in the tray Quit from the system tray, then reopen
Hook never fires Wrong event, matcher case, or untrusted folder /hooks; matchers are case-sensitive; accept the trust prompt
Hook shows an error notice Script exits non-zero or cannot start Pipe sample JSON into it by hand and check $LASTEXITCODE
Hook JSON has no effect Extra text before the JSON, or a field at the wrong level Make the script print only JSON; check field nesting in the hooks reference
Plugin installed but its skills are missing Not reloaded, or skills/ inside .claude-plugin/ /reload-plugins; check the Errors tab; move skills/ to the plugin root
Auto mode blocks routine work Classifier lacks context /permissions > Recently denied, press r to retry with approval; or use Manual mode for that task
Claude stops following earlier instructions Long session compacted them away Start fresh; move lasting instructions into CLAUDE.md or Project instructions
Answers slow down or lose track Context nearly full /compact, /clear, or a new chat

SECTION 19

Four-week learning plan

This plan takes you from the everyday claude.ai features to a guarded Claude Code project with your own skill, subagent, hook and MCP server. Each week ends with a check you can verify.

Week 1: claude.ai, no terminal

Week 2: terminal and install

Week 3: Claude Code core

Week 4: connect and automate

Further learning from Anthropic

SECTION 20

Quick reference

The terms, file locations and commands from this guide on one page.

Glossary

Term Meaning Section
Agent Model plus tools plus a loop of act, observe, decide 2
Agent-ready app An app that can be driven by a CLI, a JSON channel or an MCP server 15
Artifact A page Claude publishes to your claude.ai gallery, private until shared 6
Auto memory Notes Claude Code writes itself, per project 10
Auto mode Permission mode where a second model reviews risky actions 10
CLAUDE.md Standing instructions file Claude Code reads every session 10
CLI A program controlled by typed commands 4
Compaction Summarising older conversation to free context 2
Connector claude.ai's name for an MCP server it connects to for you 7
Context window Everything the model can see on one turn 2
Deny rule Settings entry that blocks a tool or path in every mode 10
Environment (surface) Where the agent runs and what it can touch 3
Exit code Number a program returns: 0 success; 2 blocks in hooks 4, 13
Fork Subagent that inherits the whole conversation 12
Git Version history for a folder; each saved snapshot is a commit 4
Frontmatter YAML settings block between --- lines at the top of a file 5, 8
Headless (-p) Running Claude Code once from a command, no session 15
Hook Command run automatically at a lifecycle event 13
JSON Text format for configs and messages 5
Marketplace Catalogue of plugins 14
MCP Open standard connecting AI clients to tool servers 11
MCP Inspector Tool for testing an MCP server without AI 11
Model The language model itself 2
PATH List of folders Windows searches for programs 4
Permission mode How much Claude may do without asking 10
Plan mode Claude researches and proposes before changing anything 10
Plugin Package of skills, subagents, hooks and MCP servers 14
Project (claude.ai) Workspace with its own files and instructions 6
Prompt injection Hidden instructions in content an agent reads 17
Research claude.ai mode that runs many searches and tool calls before answering 7
Scope Who a setting, server or plugin applies to: user, project, local 10, 11, 14
Skill Folder with a SKILL.md of instructions for a repeated task 8
Slash command /name typed to run a command or skill 10
stdio / HTTP transport Local process versus remote server for MCP 11
Subagent Separate Claude with its own context that returns a summary 12
Token Unit the model counts text in (about 3 to 4 characters of English, depending on the model) 2
Tool One named action the model can request 2

File locations on Windows

Type these into File Explorer's address bar. In PowerShell, write $HOME in place of %USERPROFILE% and $env:APPDATA in place of %APPDATA%.

What Where
User settings %USERPROFILE%\.claude\settings.json
Project settings (shared, local) <project>\.claude\settings.json, <project>\.claude\settings.local.json
User CLAUDE.md %USERPROFILE%\.claude\CLAUDE.md
Project CLAUDE.md <project>\CLAUDE.md or <project>\.claude\CLAUDE.md
Personal skills %USERPROFILE%\.claude\skills\<name>\SKILL.md
Project skills <project>\.claude\skills\<name>\SKILL.md
Personal subagents %USERPROFILE%\.claude\agents\<name>.md
Project subagents <project>\.claude\agents\<name>.md
MCP servers, local and user scope %USERPROFILE%\.claude.json
MCP servers, project scope <project>\.mcp.json
Auto memory %USERPROFILE%\.claude\projects\<project>\memory\
Claude desktop app MCP config %APPDATA%\Claude\claude_desktop_config.json
Claude Code program %USERPROFILE%\.local\bin\claude.exe

PowerShell essentials

POWERSHELL · C-72
Get-Location; Get-ChildItem; cd <path>; cd ..; mkdir <name>
Get-Content <file>; Set-Content <file> "<text>"; Copy-Item <a> <b>
$LASTEXITCODE
$env:PATH -split ';'
winget install --id <id> -e
<program> --help

Claude Code from the terminal

POWERSHELL · C-73
claude                                  # start a session here
claude --permission-mode manual         # ask before every action
claude -c                               # continue last conversation
claude -r                               # pick a conversation to resume
claude -p "<task>" --allowedTools "Read" # one-shot run
claude --version; claude doctor; claude update
claude mcp add <name> -- <command> [args]
claude mcp list; claude mcp get <name>; claude mcp remove <name>
claude plugin install <plugin>@<marketplace> --scope user
claude plugin list; claude plugin validate <path>
claude --plugin-dir <path>              # load a plugin for one session

Inside a Claude Code session

Keys or command Does
Esc Interrupt
Shift + Tab Cycle permission modes
@<file> Attach a file
/plan <task> Plan before changing anything
/init, /memory Create and edit CLAUDE.md
/context, /compact, /clear Manage context
/rewind Undo to a checkpoint
/mcp, /hooks, /plugin, /skills Inspect add-ons
/permissions Allow, ask and deny rules
/usage, /status, /doctor Usage, configuration, checkup
/tasks Background work, including subagents
/exit Leave