agent-ste
Use when rewriting user or planner instructions before planning or execution; when prompts contain vague verbs (optimize, improve, better, appropriate, reasonable, quickly, carefully, as needed, if necessary); when multiple agents or LLMs must interpret the same task reproducibly; when an instruction lacks explicit actor, inputs, outputs, success criteria, failure conditions, ordering, or scope; or when preparing handoffs between orchestrators and workers.
- License 许可
- MIT
- 所属仓库
- dasobral/skills
- 最近更新
- 2026-08-15
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- agent-ste
- license
- MIT
- metadata
- { "ingested_by": "skills-export", "source_platform": "portable" }
- description
- Use when rewriting user or planner instructions before planning or execution; when prompts contain vague verbs (optimize, improve, better, appropriate, reasonable, quickly, carefully, as needed, if necessary); when multiple agents or LLMs must interpret the same task reproducibly; when an instruction lacks explicit actor, inputs, outputs, success criteria, failure conditions, ordering, or scope; or when preparing handoffs between orchestrators and workers.
02 资源文件 6
- 参考文档references/checklist.md3.8 KB
- 参考文档references/examples.md10.5 KB
- 参考文档references/principles.md4.5 KB
- 参考文档references/rules.md6.2 KB
- 参考文档references/spec-v1.md8.9 KB
- 参考文档templates/instruction-block.md2.0 KB
03 SKILL.md 正文
Agent STE (Agent Simplified Technical English)
Rewrite instructions into a controlled natural language optimized for LLM execution: explicit, measurable, ordered, and reproducible across models.
Core principle: Prefer deterministic execution over natural prose. Explicitness beats brevity. Unambiguous wording beats elegant wording.
Inspiration: Principles behind ASD-STE100 (controlled language for clarity).
This skill is not ASD-STE100 and does not claim compliance. See
references/principles.md.
When to Use
- Before planning or executing non-trivial work
- Before dispatching subagents or cross-model handoffs
- When the prompt contains forbidden vague language (see below)
- When success would be judged differently by two competent agents
When NOT to use
- Pure chitchat or one-line factual Q&A with no side effects
- Already-formal specs (OpenAPI, JSON Schema, RFC text) that need no rewrite
- Emergency stop / safety refusal wording that must stay short and human
Transformation Algorithm
Execute these steps in order. Do not skip steps under time pressure.
- Objective — State exactly one primary deliverable in one sentence. Do not
join independent goals with “and” inside
OBJECTIVE. - Entities — List every actor, system, file, service, and artifact. Assign
a stable identifier (
SVC-API,FILE-auth.ts,AGENT-worker). - Glossary — For each critical term, fix one meaning. Ban synonyms for those terms inside the instruction.
- Demote vagueness — Replace every forbidden pattern with a measurable or
observable statement (
references/rules.md). - Split actions — One action per sentence. One objective per instruction block. Use a numbered list for ordering.
- Fill slots — Complete the template in
templates/instruction-block.md(actor, object, inputs, outputs, assumptions, constraints, preconditions, postconditions, invariants, success, failure, rollback, idempotency, side effects, permissions, dependencies, artifacts, validation, checklist). - Scope fence — State what is in scope and what is out of scope.
- Self-check — Answer every item in
references/checklist.mdwith YES/NO. If any required item is NO, revise before planning or execution.
Mandatory Rules (summary)
Full normative list: references/rules.md.
| Rule | Why |
|---|---|
| One action per sentence | Prevents partial execution and skipped conjuncts |
| One objective per instruction block | Stops silent scope expansion |
| One meaning per critical word | Removes synonym drift across LLMs |
| Explicit actor + object | No orphan imperatives |
| Explicit inputs + outputs | Makes dataflow checkable |
| Explicit success + failure | Defines done vs retry vs abort |
| Explicit ordering | 1 → 2 → 3 or mark PARALLEL |
| Explicit units + identifiers | No “soon”, “large”, “the file” |
| No pronouns when ≥2 entities | “It/they/this” is an entity bug |
| No forbidden vague lexicon | Forces observables |
Forbidden Patterns (detect and rewrite)
optimize · improve · better · appropriate · reasonable · quickly ·
carefully · if necessary · as needed · etc. · and so on · normally ·
probably · maybe · somehow · tidy · harden · fix up ·
make sure (without a checkable predicate) · satisfied · obvious
Rewrite table and agent-oriented slots: references/rules.md.
Output Contract
Emit an Agent STE block using templates/instruction-block.md. Then either:
- execute that block, or
- hand it to the next agent unchanged
Do not plan from the original vague prompt once an Agent STE rewrite exists.
Examples and Spec
- Worked rewrites (coding, DevOps, docs, Git, research, debug, infra, security):
references/examples.md - Objective compliance checklist:
references/checklist.md - Future machine-checkable spec (MUST/SHOULD/MAY, JSON/YAML, linting, MCP):
references/spec-v1.md
Limitations
- Does not replace domain expertise or formal verification
- Cannot invent missing requirements; mark
UNKNOWNand halt or ask - Numeric targets require a source (user, SLO, benchmark); else state
METRIC-SOURCE: UNKNOWNand do not fabricate thresholds - Over-application to trivial chat wastes tokens — respect When NOT to use
- Not a license to ignore higher-priority safety or repository rules
Rationalizations — Do Not Skip
| Excuse | Reality |
|---|---|
| "Too simple to rewrite" | Simple vague prompts still fork across models |
| "No time for slots" | Missing success criteria causes expensive rework |
| "Context implies the metric" | Another LLM does not share your implied context |
| "I'll keep optimize but add detail" | Forbidden lexicon remains non-deterministic |
| "Pronouns are fine; there are only two things" | Two things is exactly when pronouns collide |
| "Rollback is overkill for a docs tweak" | State ROLLBACK: none — read-only explicitly |
| "One OBJECTIVE can list three goals with and" | That is three blocks or gated phases — split them |
Red flags — STOP and rewrite: residual etc.; success = “looks good”;
actor missing; two actions joined by “and”; synonym pairs for the same concept;
if necessary without a boolean predicate.
04 同仓库更多 Skills
fixture-skill
Fixture skill used to prove trust inventory coverage.
write-conformant-code
Write code in any language that matches project conventions exactly. Reads CODING_REQUIREMENTS.md, prior review output, or built-in defaults before writing. Part of Codecraft pipeline: analyze → write → review → audit. Use when implementing features, functions, or any code that should fit seamlessly into an existing codebase.
test-crypto-interoperability
Record reproducible cryptographic interoperability cases, including negotiation, fallback, sizes, latency, fragmentation, and negative tests.
scientific-platform-architect
Design scientific computing platforms combining HPC kernels with modern infrastructure. Enforces compute-layer separation (Rust/PyTorch/Python), emulator-first thinking, and right-sized infrastructure. Use for MCMC pipelines, neural emulators, platform design, or scientific code review.
rust-systems
Rust systems engineering for daemons, HTTP APIs, and AOS components. Enforces Tokio/Axum async-first, LiteLLM-compatible OpenAI APIs, workspace layouts, tracing, and systemd-compatible daemon design. Part of AOS Stack plugin. Use for .rs files, Cargo.toml, daemons, or Rust orchestrator services.
run-agent-attack-replay
Aggregate isolated agent attack trials with deterministic side-effect assertions, Wilson intervals, and preserved trace evidence.