pro-workflow is a set of hooks, skills, and agents distilled from how power users ship. This page follows one session from start to finish and shows where each part runs.
session-start loads recent learnings from the SQLite store and shows session stats. setup-hook detects the project type.
CLAUDE.md | Entry point, imports the rest |
AGENTS.md | Workflow rules |
SOUL.md | Style preferences |
LEARNED.md | Filled in from your corrections |
| Dev | Code first, iterate quickly |
| Review | Security focus, read-only audit |
| Research | Explore broadly, summarize findings |
prompt-submit spots corrections like “don’t do that again” and suggests /learn, and loads relevant wiki pages. drift-detector restates your original goal when, after 6 or more edits, a new prompt looks unrelated to it.
[LEARN] ruleLEARNED.md and the databaseTurn on Laya (local) or Jev (hosted) to catch corrections the keyword check misses, in tens of milliseconds. Off by default. It only suggests /learn; it never allows or denies anything.
Research wikis in the same SQLite store with full-text search, built by wiki-builder and grown by a budget-capped research loop over web, arXiv, and GitHub.
The only stop that can block. Every blocker in v3.4 parses the real command first, so heredocs and quoted text no longer trip it.
| When Claude calls | Script | What happens | Effect |
|---|---|---|---|
Bash | git-blast-radius | Stops destructive git: force-pushes, hard resets, branch deletes, history rewrites | blocks |
Bash | commit-validate | Requires a conventional commit message on a real git commit | blocks |
Edit Write | secret-scan | Refuses to write keys, tokens, or secret-like files | blocks |
Read | reread-tracker | Tells Claude it already read an unchanged file. Blocking is opt-in | warns |
Edit Write | read-before-write | Flags edits to files Claude has not read | warns |
Edit Write | quality-gate | Calls a review checkpoint every few edits | warns |
Bash | pre-commit-check pre-push-check | Reminds Claude to lint, typecheck, and test first | warns |
| any tool | tool-call-budget | Suggests wrapping up as tool calls pile up | warns |
Each prompt and denial is logged. permission-analyst and /permission-tuner read those logs with the transcript’s toolDenialKind and suggest the smallest fix: a narrower ask rule, a hook repair, or an autoMode entry.
post-edit-check flags leftover console.log calls and other issues in edited files. test-failure-check turns a failing test into a learning. reread-tracker records each completed read, and tool-failure tracks errors.
Lifecycle hooks log each agent and task, check task quality, and flag idle teammates.
Breaks work into steps. Reads and proposes, never edits.
Logic, edge cases, errors, security, performance, tests.
Scores readiness 0 to 100 in an isolated worktree.
Research, plan, implement, review, with /develop.
Reproduces, finds the root cause, verifies the fix.
Audits CLAUDE.md size, skill overhead, MCP tool count.
Turns denials into narrow rules and autoMode entries.
Token use, cache hits, model choice.
Scout scores five areas, 20 points each:
pre-compact saves edit and prompt counts and the current goal; post-compact restores them from the summary.
| Context rule | Why |
|---|---|
| Read before you edit | Understand the code first |
| Compact at task boundaries | Free the token budget cleanly |
| Fewer than 10 MCP servers, 80 tools | Less noise in every request |
| Use subagents for reading-heavy work | Protect the main context |
| Keep CLAUDE.md under 150 lines | Avoid instruction drift |
learn-capture saves any [LEARN] block Claude wrote to the database. session-check adds reminders based on what happened. stop-failure names the API error when a turn fails: rate limit, overload, auth, or server error.
session-end saves the session’s stats. The /wrap-up ritual closes the loop:
git status, uncommitted work, TODOsconfig-watchercwd-changedcdfile-changed.env, package.json, and other config filesnotification-handlerSkills you call on demand, and the habits the hooks reinforce.
claude -w| Task | Model | Effort |
|---|---|---|
| Quick fixes, read-only exploration | Haiku 4.5 | low |
| Features | Sonnet 5 | high |
| Refactors and architecture | Opus 5.5 | xhigh |
| Long-horizon work | Fable 5.1 | high |
Also in the library: skill-router, tdd, domain-modeling, design-engineering, writing-guidelines, improve-architecture, context-engineering, batch-orchestration, thoroughness-scoring, sprint-status, agent-teams, safe-mode, token-efficiency, skill-optimizer, bug-capture, module-map, and more. 41 skills in all.
| Self-correction | Claude learns from your corrections |
| Parallel worktrees | No waiting on builds or tests |
| Wrap-up ritual | End each session with intention |
| Split memory | Modular CLAUDE.md for big projects |
| 80/20 review | Review at checkpoints, not every change |
| Model selection | Route by task and effort |
| Context discipline | Spend the context budget on purpose |
| Learning log | Write down what each session taught |
Each phase has to pass its gate before the next starts. Commands, agents, and skills chain the same way: a command starts an agent, the agent runs skills, and agents hand off through Task, never directly.
“80% of my code is written by AI, 20% is spent reviewing and correcting it.”