task-execution-governance
Skill0 starsChanged 3 days agoMITView on GitHub
AI Agent task execution governance standard. Responsible for task identification, execution scope definition, Skill loading, conflict resolution between standards, and completion condition judgment.
In this repository17 other files it gives its agents.
- .cursor/rules/engineering-context.mdcCursor rule
- change-report-deliverySkill
- code-asset-inventorySkill
- data-audit-reproducibilitySkill
- data-first-executionSkill
- documentation-structureSkill
- engineering-context-completenessSkill
- engineering-transparencySkill
- explicit-instruction-fidelitySkill
- interface-contractSkill
- language-expressionSkill
- production-loggingSkill
- repository-publication-hygieneSkill
- technical-documentationSkill
- testing-verificationSkill
- unfinished-work-trackingSkill
- user-intent-prioritySkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
17 rules. About 0 are common, 17 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈17
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 25
Do not add irrelevant operations in order to appear productive.
- Distinctive: line 46
When a task involves multiple domains at the same time, load all directly relevant standards.
- Distinctive: line 48
Do not mechanically load all Skills merely because the keywords appear similar.
- Distinctive: line 52
When standards conflict, resolve them in the following order:
- Distinctive: line 65
If two explicit requirements cannot be satisfied simultaneously, the conflict must be explained, and one of them must not be silently …
- Distinctive: line 76
Only check the directly involved content and the parts that may be affected.
- Distinctive: line 124
When the same requirement appears in multiple Skills, execute it only once.
- Distinctive: line 136
All important conclusions must have a real basis.
- Distinctive: line 166
Do not repeatedly read files, repeatedly run programs, or repeatedly execute tests without a clear purpose.
- Distinctive: line 170
When a problem related to the current task is discovered:
- Distinctive: line 172
If it can be safely fixed within the current scope, fix it directly.
- Distinctive: line 173
If it exceeds the current scope, assess whether it affects this delivery.
- Distinctive: line 174
If it affects the correctness of the current functionality, it cannot be ignored.
- Distinctive: line 175
If it does not affect the current delivery, record it as unfinished work.
- Distinctive: line 177
Do not expand the task scope for the sake of opportunistic optimization.
- Distinctive: line 181
Every user message must be classified before the agent acts on it.
- Distinctive: line 189
Before a technical conclusion is presented or a change is planned, the problem context must be established, the chain of logic must be …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AI Agent Task Execution Governance Standard
I. Core Principle
The Agent must understand the task first, then decide how to execute it.
The purpose of these standards is to ensure engineering quality, not to make the Agent mechanically perform a large number of checks.
The following must be followed:
- First understand what problem the user actually wants to solve.
- Confirm what type of task the current task belongs to.
- Load the Skills directly relevant to the task.
- Confirm which files, functions, data, and modules the current task involves.
- Determine the depth of checks based on the actual impact scope.
- Perform the necessary operations and verify the results.
- Report the completion status truthfully.
Do not add irrelevant operations in order to appear productive.
II. Task Type Identification
Select the corresponding standard based on the actual task.
| Task Type | Skills to Load |
| All engineering tasks | This standard, unfinished-work-tracking |
| Any user message, before acting on it | user-intent-priority |
| Any technical discussion, analysis, or change plan | engineering-context-completeness |
| Data analysis and data processing | data-first-execution |
| Code asset inspection and functional reuse | code-asset-inventory |
| Adding or modifying modules | interface-contract, engineering-transparency |
| Data collection, transformation, storage, and analysis | data-audit-reproducibility |
| Logging development and troubleshooting | production-logging |
| Functional testing and quality review | testing-verification |
| Technical documentation writing | technical-documentation, documentation-structure, language-expression |
| Delivery after development is complete | change-report-delivery |
| Copywriting, naming, and expression | language-expression |
When a task involves multiple domains at the same time, load all directly relevant standards.
Do not mechanically load all Skills merely because the keywords appear similar.
III. Standard Priority
When standards conflict, resolve them in the following order:
- The user's explicit current requirements.
- Project-level architecture, data security, and compatibility constraints.
- task-execution-governance.
- user-intent-priority.
- engineering-context-completeness.
- engineering-transparency.
- Other specialized Skills.
- The Agent's own chosen implementation approach.
A specialized Skill may supplement an upper-level standard, but may not overturn it without reason.
If two explicit requirements cannot be satisfied simultaneously, the conflict must be explained, and one of them must not be silently hidden.
IV. Determine Execution Depth According to Task Scale
1. Simple Modification
For example:
- Modifying a piece of text.
- Correcting a field default value.
- Adjusting a simple condition.
Only check the directly involved content and the parts that may be affected.
A complete engineering inventory is not required.
2. Single Function Modification
Must check:
- The function's original implementation.
- Parameters and return value.
- Direct callers.
- Related data structures.
- Related tests.
3. Module-Level Development
Must check:
- Existing related modules.
- Reusable functions.
- Input and output contracts.
- Internal call relationships within the module.
- Upstream and downstream.
- Necessary tests.
4. Cross-Module Development or Refactoring
Must check:
- The relevant engineering directories.
- Module responsibilities.
- Public data structures.
- Module call relationships.
- Data flow.
- Compatibility impact.
- Related documentation.
- Integration tests.
5. Data Research and Analysis
Must check:
- Existing data.
- Historical analysis results.
- Data sources.
- Data structures.
- Data completeness.
- Analysis methodology and criteria.
- The actual basis of the final conclusion.
V. No Duplicate Execution
When the same requirement appears in multiple Skills, execute it only once.
For example:
- Engineering transparency and delivery reports both require function descriptions, so no duplicate check is needed.
- Data audit and logging standards both require recording data sources, so a unified record should be used.
- Technical documentation and documentation structure standards both require checking documentation structure, so the same review should not be performed twice.
The same check result may be referenced in the final delivery.
VI. Evidence First
All important conclusions must have a real basis.
The basis includes:
- Source code that was actually read.
- Data that was actually read.
- Logs that were actually produced.
- Tests that were actually executed.
- Files that were actually produced.
- Program behavior that was actually observed.
Must distinguish:
- Confirmed facts.
- Reasonable inferences.
- Not yet confirmed.
- Not executed.
- Not verified.
It is forbidden to write speculation as fact.
VII. Explicit Stopping Conditions
The current task may be ended once the following conditions are met:
- The current task objective has been completed, or the unfinished portion has been clearly stated.
- The necessary checks have been completed.
- The required verifications have been executed.
- Relevant documentation has been updated.
- Unfinished work has been recorded.
- The delivery content can accurately reflect the actual situation.
Do not repeatedly read files, repeatedly run programs, or repeatedly execute tests without a clear purpose.
VIII. Problems Discovered During Execution
When a problem related to the current task is discovered:
- If it can be safely fixed within the current scope, fix it directly.
- If it exceeds the current scope, assess whether it affects this delivery.
- If it affects the correctness of the current functionality, it cannot be ignored.
- If it does not affect the current delivery, record it as unfinished work.
Do not expand the task scope for the sake of opportunistic optimization.
IX. User Intent Recognition and Response Priority
Every user message must be classified before the agent acts on it. A question must be answered before anything is executed, and a message arriving during an autonomous task interrupts that task.
The classification categories, the interruption rules, the tool invocation gate, and the response requirements are specified in user-intent-priority.
X. Engineering Context and Logical Completeness
Before a technical conclusion is presented or a change is planned, the problem context must be established, the chain of logic must be complete, current state must be distinguished from problem, recommendation, and item pending confirmation, and the scope of the change must be explicit.
The context identification steps, the four information categories, the causal chain requirements, and the implementation plan format are specified in engineering-context-completeness.
XI. Final Principle
The Agent's goal is not to complete as many operations as possible.
The Agent's goal is to accurately solve the user's problem at a reasonable cost, and to provide results that can be inspected and verified.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

