AGENT SKILL

ultrawork

Maximum-intensity execution protocol. Activated when [ULTRAWORK ACTIVE] header is present in delegation prompt. Enforces pre-implementation gates, zero-excuses completion, anti-rationalization, and mandatory verification.

License 许可
Apache-2.0
所属仓库
yqecea/gem-swarm
最近更新
2026-04-26
Stars
★ 1
Forks
1

01 前言 Front Matter

name
ultrawork
version
1
priority
CRITICAL
description
Maximum-intensity execution protocol. Activated when [ULTRAWORK ACTIVE] header is present in delegation prompt. Enforces pre-implementation gates, zero-excuses completion, anti-rationalization, and mandatory verification.
allowed-tools
Read, Write, Edit, Glob, Grep, Shell

03 SKILL.md 正文

Ultrawork Protocol — Agent Execution Rules

TRIGGER: This skill activates when your delegation prompt contains [ULTRAWORK ACTIVE]. It applies ON TOP of your normal agent methodology. It does NOT replace your specialization — it intensifies it.


Pre-Implementation Gate

YOU MUST NOT START ANY IMPLEMENTATION UNTIL YOU ARE 100% CERTAIN.

Before writing a single line of code, you MUST:
FULLY UNDERSTAND what the orchestrator ACTUALLY wants (not what you ASSUME)
READ every file listed in "Files to modify" before changing anything
HAVE A CLEAR PLAN — if your plan is vague, YOUR WORK WILL FAIL
RESOLVE ALL AMBIGUITY — if anything is unclear, state it in your Task Report

Signs you are NOT ready to implement:

  • You're making assumptions about file structure
  • You're unsure which files exist
  • You don't understand how existing code works
  • Your approach has "probably" or "maybe" in it

Zero Excuses

The orchestrator's delegation is sacred. You MUST fulfill it exactly.

VIOLATION CONSEQUENCE
"I couldn't because..." UNACCEPTABLE. Find a way or report a blocker.
"This is a simplified version..." UNACCEPTABLE. Deliver FULL implementation.
"You can extend this later..." UNACCEPTABLE. Finish it NOW.
"Due to limitations..." UNACCEPTABLE. Use all available tools.
"I made some assumptions..." UNACCEPTABLE. State them as explicit questions in Task Report.

Anti-Rationalization Shield

If you catch yourself forming any of these phrases, STOP and verify.

If you're about to say... It means... Do this instead
"This looks correct" You read but didn't verify Run the build. Check the output.
"This is probably fine" You're uncertain State what's uncertain. Test it.
"I believe this should work" You haven't confirmed Run it. Then say "Verified: it works."
"No issues found" You stopped looking List what you checked and what you didn't

The shield is working if: your Task Report contains specific evidence (command output, test results), not general assessments.


Mandatory Verification Gates

Before reporting your phase as complete, you MUST pass ALL applicable gates:

Gate What to do
Build Run the build command. It must succeed.
Test Run tests. They must pass.
Output Verify the output contains expected data. Check it.
Integration Verify your changes don't break adjacent files.

Do NOT report success based on "I believe this should work." Run it. Confirm it. Then report.


Manual QA Mandate

Type checks catch type errors, NOT functional bugs. Your work is NOT verified until you MANUALLY test it.

If your change... YOU MUST...
Creates new files Verify the files exist and have correct content
Adds a component Verify it renders (describe expected output)
Adds a function Call it with test inputs, show the output
Modifies config Load the config, verify it parses
Adds CLI behavior Run the command, show the output
Changes build output Run the build, verify output files

Anti-AI-Filler

During Ultrawork, enforce strict output hygiene:

Rule Enforcement
No placeholder comments Zero // TODO, // implement here, // add more
No type suppression Zero @ts-ignore, as any, eslint-disable
No empty error handling Zero catch(e) {} — always handle errors
No catch-all files Zero utils.ts, helpers.ts with single functions
Comments explain WHY Never explain WHAT the code does — it's self-documenting
No AI filler phrases Zero "Here's the implementation", "Let's create" in code comments

No Partial Completion

Your phase deliverables are a contract. Every item must be delivered.

  • Every file listed in "Files to create" MUST be created with full implementation
  • Every file listed in "Files to modify" MUST be modified as specified
  • Every deliverable listed MUST be present in your output
  • No stubs: if a function exists, it has a real implementation
  • No skipped states: if a component renders, ALL states work (loading, error, empty, success)

The phase is NOT done until:

  1. All deliverables are implemented
  2. Code compiles without errors
  3. Tests pass (or were written and pass)
  4. You have verified the output, not just the process

Parallel Awareness

When running in Ultrawork parallel mode, you may be one of several instances of the same agent running simultaneously. Follow these rules:

  1. Own your files: Only modify files explicitly listed in YOUR delegation prompt
  2. Don't touch shared files: Layout files, barrel exports, and config files belong to the integration phase
  3. Export cleanly: Make your components importable without side effects
  4. No assumptions about siblings: Other parallel instances produce their own deliverables — don't reference them unless explicitly told to
  5. Self-contained validation: Run validation only on YOUR files

Task Report Enhancement

In Ultrawork mode, your Task Report MUST include:

## Task Report

Status: [success|failure|partial]

### Verification Evidence
- Build: [command run] → [pass/fail]
- Test: [command run] → [pass/fail]
- Manual QA: [what was verified] → [result]

### Files Created
- [path]: [purpose]

### Files Modified
- [path]: [what changed]

### Anti-Slop Self-Check
- [ ] Zero TODO/placeholder tokens
- [ ] Zero @ts-ignore/as any
- [ ] Zero empty catch blocks
- [ ] All component states handled
- [ ] All deliverables present

## Downstream Context
[context for next phase]

04 同仓库更多 Skills