Skip to content
K Koda Intelligence
Subscribe
All Dojos
AI Foundations Claude Code Cowork M365 Copilot NotebookLM ChatGPT Lovable Perplexity
Discipline 01 · 3 white, 8 brown, 4 black

Claude Code Dojo

Install it, teach it your codebase, then run a fleet of agents. The capstone is a self-improving loop you can hand over.

Your rank

Untied

15 modules · 174 prompts · 117 sourced
White · 3 modulesBrown · 8 modulesBlack · 4 modules
White Belt3 modules46 promptsNot started
Brown Belt8 modules79 promptsNot started
Black Belt4 modules49 promptsNot started
Version 2.0.11·Revised 2026-09-13·Refreshed weekly·55 sources cited
Inventory
Modules15
Prompts174
Prompts sourced117
Sources cited55
Your session
Modules woven0
Prompts ticked0
Minutes trained (est)0
Start training
// 01 / The ladder

Fifteen modules

White belt is a workspace that boots with context. Black belt is a loop that improves itself and records the movement.

White Belt
Fundamentals
Not started
01Up next
Getting Started: Setup and Interface
Sessions, CLI invocation patterns, slash commands and the keyboard moves that keep an agent on course, taking you from fresh install to fluent terminal operator.
~45 min (est)16 prompts 16 beginner
02
The Project Brain: CLAUDE.md Mastery
A CLAUDE.md system that makes Claude measurably smarter every time you use it, so you stop re-explaining your project every session.
~60 min (est)22 prompts 17 beginner · 4 intermediate · 1 advanced
03
Permission Modes and Plan Mode
The right autonomy level for each job, with plan mode as your default safety net, instead of approving every single action by hand.
~45 min (est)8 prompts 5 beginner · 2 intermediate · 1 advanced
Capstone · White BeltKnot open
The Bulletproof Workspace
Prove you can build a workspace that makes every future session smarter than the last. You will set up a project where Claude Code boots with full context, safe permissions, and a live tool connection, then prove it works by cold-starting a fresh session that needs zero re-explaining. This is the foundation every other belt stands on.
  • Run /init in a fresh project and expand the generated CLAUDE.md to cover project knowledge, your working preferences, key commands, and a decision log section.
  • Add a scoped CLAUDE.md inside at least one subdirectory with rules that apply only to work in that folder, and confirm Claude follows them when editing files there.
  • Connect one MCP server with claude mcp add and verify it responds by checking /mcp and running a real query through it.
  • Audit your setup with /permissions, tighten it to the lowest settings that still let you work, and record what you changed and why in CLAUDE.md.
  • Teach Claude one durable preference mid-session using the # memory shortcut and confirm it was written to the right file.
  • Cold-start test: open a brand-new session and ask a project-specific question; pass only if Claude answers correctly with zero re-explaining from you.
DeliverableA committed workspace repo containing a root CLAUDE.md, at least one scoped subdirectory CLAUDE.md, a verified MCP server connection, and a written permissions audit that a teammate could adopt in ten minutes.
Brown Belt
Advanced
Not started
04
Context Management and Compaction
Long productive runs where you measure, budget and preserve context on purpose, instead of sessions that quietly degrade into confusion.
~55 min (est)14 prompts 2 beginner · 5 intermediate · 7 advanced
05
Skills and Automation
A library of skills, slash commands and hooks that run your workflows on demand, so you stop re-typing your best prompts every session.
~50 min (est)9 prompts 1 beginner · 4 intermediate · 4 advanced
06
MCP Servers and Plugins
A connected command center: add, secure and chain MCP servers so one prompt can span your browser, inbox, GitHub and internal tools, starting from a stock Claude Code install.
~55 min (est)11 prompts 1 beginner · 5 intermediate · 5 advanced
07
Agent Harnesses
The machine around the model, audited and upgraded, so the same LLM completes dramatically more work without you, instead of you blaming the model.
~45 min (est)8 prompts 1 beginner · 4 intermediate · 3 advanced
10
Browser Automation
An escalation ladder that replaces copy-pasting web content by hand: HTTP APIs first, HTML parsing second, full browser control only when the job demands it.
~45 min (est)9 prompts 2 beginner · 4 intermediate · 3 advanced
11
Security
Proof rather than hope: systematic audits for secrets, hallucinated packages, vulnerable dependencies, exposed database tables, and the new attack surfaces AI tooling itself creates.
~50 min (est)10 prompts 5 intermediate · 5 advanced
13
Common Mistakes and Anti-patterns
The operator loop, in place of the nine mistakes that silently cap everyone's results: scoped asks, verified output, corrected behavior, codified lessons.
~50 min (est)9 prompts 2 beginner · 6 intermediate · 1 advanced
15
Hooks Recipe Book
Nine working hook recipes you can paste into settings.json today, from activity logging to commit-blocking security gates, so your rules are enforced mechanically instead of remembered.
~55 min (est)9 prompts 2 beginner · 5 intermediate · 2 advanced
Capstone · Brown BeltKnot open
The Research Machine
Build a research system, not a research session. You will create a reusable skill that fans a question out to three subagents, forces their findings through a consensus filter so no single agent's hallucination survives, and ships a source-linked report. The pass bar is reuse: the same skill must work on a second topic without editing.
  • Write a SKILL.md defining a reusable research skill with a clear trigger phrase, required inputs, and a fixed output format.
  • Have the skill fan out three subagents in parallel, each assigned a distinct angle on the same question: supporting evidence, contradicting evidence, and prior art or market context.
  • Merge the three findings with an explicit consensus rule: a claim only enters the final report if at least two subagents independently support it, and dropped claims are listed with the reason.
  • Ship the merged output as one structured report with an executive summary, a confidence label on every claim, and a working source link for every fact.
  • Rerun the skill on a completely different topic without editing SKILL.md, proving the system generalizes beyond its first use.
DeliverableA reusable research skill plus two published structured reports it produced on different topics, each with consensus-filtered, source-linked claims.
Black Belt
Mastery
Not started
08
Sub-agents and Agent Teams
A coordinated fleet rather than one Claude in one window: parallel fan-out, consensus reviews, adversarial debates, and pipelines with quality gates that actually reject.
~55 min (est)14 prompts 4 intermediate · 10 advanced
09
Auto Research Loops
Self-driving optimization instead of hand-tuning: measure-change-remeasure loops that grind a metric toward a target while you do something else.
~50 min (est)10 prompts 4 intermediate · 6 advanced
12
Workspace Org, Deployment, and the Future
An organized multi-workspace operation in place of one messy folder and one session at a time: parallel agents in worktrees, groomed memory, headless pipelines, and a deliberate roadmap for what to automate next.
~50 min (est)10 prompts 4 intermediate · 6 advanced
14
Power User Workflows
Full agentic loops, disciplined handoffs, headless pipelines and parallel sessions: operating Claude Code rather than prompting it one request at a time.
~60 min (est)15 prompts 4 intermediate · 11 advanced
Capstone · Black BeltKnot open
The Self-Improving Loop
Build a system that makes a codebase measurably better without you in the loop. You will script an autonomous cycle that measures a real metric, proposes one change, applies it, re-measures, and either keeps or reverts the result, with every iteration committed and auditable. This is the belt where you stop operating Claude Code and start engineering with it.
  • Pick one measurable metric (test coverage, Lighthouse score, or build time), record a baseline, and save the exact command that produced it to the repo.
  • Build a loop script that calls headless claude -p each iteration to analyze the current state, propose exactly one change, and apply it.
  • Add hard stop conditions to the loop: a target metric value and a maximum attempt count, whichever hits first.
  • Make every iteration auditable by committing each change with the before and after metric values in the commit message.
  • Add a safety rail that automatically reverts any iteration whose re-measurement shows the metric got worse, before the next pass begins.
  • Run the loop to completion and write a one-page postmortem covering total metric movement, cost in time and tokens, and the single change that mattered most.
DeliverableA repo containing the loop script, a git history where every commit records measured metric movement, and a one-page postmortem of the run.
Beginner Intermediate Advanced Bar length is the module's prompt count against the longest module here; its segments are that module's own mix. Minutes are the curriculum's per-module estimates, summed over what you have woven.
One cell is one module, at its belt cloth Woven cell: module complete Knot open until the belt's last module is woven. Kept on this device only.
// SOURCES

Curriculum Record

Every published source this curriculum cites: 55 distinct references, counted from the curriculum brief and from the prompts themselves. Prompts that quote one carry a source line; the rest are marked unsourced author prompts.

  1. 01code.claude.comskills
  2. 02code.claude.comcommon-workflows
  3. 03code.claude.comsub-agents
  4. 04code.claude.comsettings
  5. 05code.claude.comsecurity
  6. 06docs.anthropic.comclaude-code/quickstart
  7. 07docs.anthropic.comclaude-code/overview
  8. 08docs.anthropic.comcli-reference
  9. 09support.claude.comclaude-code-cheatsheet
  10. 10code.claude.comcommunications-kit
  11. 11code.claude.commemory
  12. 12claudify.techclaude-code-best-practices
  13. 13agensi.ioclaude-code-best-practices-2026
  14. 14builder.ioclaude-code-tips-best-practices
  15. 15institute.sfeir.combest-practices
  16. 16hannahstulberg.substack.comclaude-code-for-everything-30-claude-code-tips-and-tricks
  17. 17ranthebuilder.cloudclaude-code-best-practices-lessons-from-real-projects
  18. 18code.claude.commemory
  19. 19datacamp.comclaude-code-best-practices
  20. 20marmelab.comclaude-code-tips-i-wish-id-had-from-day-one
  21. 21code.claude.combest-practices
  22. 22dev.toclaude-code-best-practices-15-tips-from-running-6-projects-2026-9eb
  23. 23code.claude.compermission-modes
  24. 24support.claude.comclaude-code-power-user-tips
  25. 25code.claude.comcontext-window
  26. 26support.claude.commodels-usage-and-limits-in-claude-code
  27. 27rosmur.github.ioclaudecode-best-practices
  28. 28dev.to10-battle-tested-claude-code-practices-4n81
  29. 29iwoszapar.comclaude-code-best-practices
  30. 30youtube.comwatch?v=EJyuu6zlQCg
  31. 31github.comclaude-code-best-practice
  32. 32superdevacademy.comclaude-ai-2026-guide-coding-tips-tricks
  33. 33code.claude.comchrome
  34. 34code.claude.commcp
  35. 35youtube.comwatch?v=T5jylUte3J8
  36. 36docs.anthropic.comclaude-code
  37. 37makerkit.devclaude-code-best-practices
  38. 38code.claude.compermissions
  39. 39code.claude.comagent-teams
  40. 40code.claude.comfeatures-overview
  41. 41code.claude.comworkflows
  42. 42code.claude.comheadless
  43. 43code.claude.comchangelog
  44. 44code.claude.comscheduled-tasks
  45. 45dev.toclaude-code-workflow-best-practices-that-ship-code-na
  46. 46code.claude.comclaude-security
  47. 47support.claude.comapi-key-best-practices-keeping-your-keys-safe-and-secure
  48. 48anthropic.comclaude-code
  49. 49github.comclaude-code-tips
  50. 50agenticcoding.substack.com32-claude-code-tips-from-basics-to
  51. 51youtube.comwatch?v=WwdIYp5fuxY
  52. 52github.comclaude-best-practice
  53. 53code.claude.comhooks-guide
  54. 54code.claude.comagent-sdk/hooks
  55. 55code.claude.comhooks
All Modules

Get the morning Signal

175 editions so far, one a day. Unsubscribe anytime.