Load Context
- Read recent git commits: git log --oneline -5
- Check TaskList for pending/in_progress tasks (if any exist)
- Read 1shot/DECISIONS.md, 1shot/BLOCKERS.md if present
- Read 1shot/LLM-OVERVIEW.md if present (quick project orientation)
Ask What's Next
```
"What are you working on?"
```
Pre-Flight Review (if the user describes a non-trivial change)
- Check providers: command -v codex >/dev/null 2>&1 && echo "codex: yes"
- If codex available, dispatch a quick adversarial take before starting
- Follow dispatch protocol (see ~/.claude/skills/_shared/dispatch.md)
- This is advisory, not a gate. Don't stall on it.
Docs Check (if the task uses any external library, API, or tool)
- Check local cache: cat ~/github/docs-cache/docs/cache/.index.md
- If the tool is missing → run /doc <name> <url> to cache it first
- Use cached docs as source of truth — do NOT rely on training data for syntax
- If no external tools involved: skip this step
Skill Discovery (if the task is specialized)
- Check 1shot/skills/ — already pulled skills for this project
- If the task involves a specific domain, tool, or API not covered by core skills:
```
./scripts/skillsmp-search.sh "<task type>" --install
```
- General tasks (write tests, fix bug, refactor): skip search, proceed
- Specialized domains (security, blockchain, ML, infra, specific parsers): search
Execute in Burn-Down Mode
- Complete one task fully before starting next
- If blocked > 2 attempts: log to 1shot/BLOCKERS.md, skip, continue
- No "pending review" — either done or blocked
- Do NOT create TaskList items for every little thing — just do the work
- After each significant change, dispatch a review via the dispatch protocol (see ~/.claude/skills/_shared/dispatch.md)
- If codex unavailable → Claude handles review inline

Show Summary on Completion

📊 Session Summary
├─ Tasks completed: X
├─ Files changed: Y
├─ Codex reviews: N (issues found: M)
├─ Skills used: [list or "core only"]
└─ Next: [next task or "all done"]

Provider Routing

See ~/.claude/skills/_shared/dispatch.md for the dispatch protocol. See ~/.claude/skills/_shared/providers.md for provider detection and commands.

Short-specific routing: Short dispatches workers for advisory reviews (pre-flight + post-completion). Uses category-based worker selection — the resolver returns workers ordered by category preference. Claude handles everything else. If workers are unavailable, zero degradation.

Scope

Optional scope limits work to matching files:

/short src/auth/*.ts    # Only work on auth files

`1shot/` Convention

Logs and state live in 1shot/, not at the project root:

1shot/DECISIONS.md — decision log
1shot/BLOCKERS.md — blocked items
1shot/skills/ — project-local SkillsMP skills

Only AGENTS.md and CLAUDE.md belong at the root.

Decision Defaults (Don't Ask)

Ambiguity	Default
Multiple implementations	Simplest
Naming	Follow existing pattern
Refactor opportunity	Skip unless blocking
Error handling	Match surrounding code
SkillsMP search?	Specialized domain → yes; general task → no
Codex review?	Always run if available (advisory only)

When truly ambiguous, pick option A, note in 1shot/DECISIONS.md.

Auto-Approved Actions

Reading any file
Writing to scope-matched files
Running tests and linters
Creating/updating any file under 1shot/
Running ./scripts/skillsmp-search.sh
Calling Codex CLI via bash
Git commit (not push)

Requires Confirmation

short

/short — Quick Iteration Operator

Usage

Behavior