agent-prompt-doctrine

星标0

分支0

更新时间2026年7月5日 16:41

Eight falsifiable rules for prompting CLI coding agents and review lanes, each derived from an observed failing-form vs working-form pair. Use when composing any prompt for a coding agent or reviewer fanout.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

frankekn

frankekn/agent-engineering-skills

打开 GitHub 仓库查看创作者相关仓库

下载

Agent Prompt Doctrine

Empirical origin: one day of orchestrated "review the whole package" fanouts burned >100M input tokens, produced four literal one-word REJECT verdicts, and shipped nothing. Weeks later, seventeen commit-sized one-shot prompts against the same codebase each ended "N passed" on a quarter of the tokens. The difference was never the model — it was the prompt shape.

The eight rules

One prompt = one verifiable diff. Whole-package verdict requests get defensive rejections; "review ONLY these 3 files" / "make EXACTLY these changes" completes cleanly.
Bound the evidence set. Inline the exact files/spec text the agent needs; lanes forced to re-read a whole repo burn tokens and return mush.
Embed the acceptance command. Name the exact test command (with interpreter path) that must pass. "It should work" prompts end in the future tense.
Enumerate prohibitions. Agents self-audit against explicit "do NOT reformat / do NOT add error handling / do NOT touch other files" lists. Unstated constraints get violated.
Scope beats format-nagging. Scolding "give a terminal verdict, not one word" half-works; narrowing scope to one diff with evidence inlined actually fixes outcomes.
No status protocols in one-shots. A "send WORKING before long passes" boilerplate caused lanes to END on the heartbeat — "WORKING: ..." as the final answer.
Never respawn an unchanged fanout. Same prompt + same scope = same failure. Change the decomposition, not the dice.
Pre-pay investigation. Implementer prompts contain conclusions (paths, line refs, decided approach), not open questions. Investigation is a separate, cheaper, read-only pass.

Prompt skeleton

<ONE commit-sized task statement>
Files: <exact paths, insertion points, line refs if known>
Spec: <verbatim spec text — do not paraphrase your plan>
Do NOT: <reformat / add defensive code / touch other files / run long test suites>
Must pass: <exact test command>

Timeouts

Full test suites hang agents (observed: a unit run stalled at 17% for 7+ minutes). Agents run focused test files only; the orchestrator runs the broad suite at the review gate.

同仓库更多 Skills

同仓库

adversarial-review

frankekn/agent-engineering-skills

Cheap-model adversarial verification pass for the end of every implementation phase and before any merge. Framing beats model tier — a low-effort model with the right question finds real bugs the expensive authoring model missed.

2026-07-050

cli-agent-driver

frankekn/agent-engineering-skills

Orchestrate CLI coding agents (codex, grok, aider, etc.) so agents write ALL implementation code while the orchestrator plans, reviews diffs, runs tests, and commits. Use for any implementation task in a multi-agent setup.

2026-07-050

measure-dont-believe

frankekn/agent-engineering-skills

Discipline for algorithm and ranking changes — hand-designed signals are usually wrong in direction, not just magnitude. Build an offline backtest first; the backtest script becomes the behavior spec; numeric gates decide mergeability.

2026-07-050

no-ai-slop

frankekn/agent-engineering-skills

The laws that keep plausible-but-broken AI-generated code away from main, derived from a real 104-PR bulk-merge disaster. Apply to every merge, every "done" claim, and every batch operation.

2026-07-050

verify-inheritance

frankekn/agent-engineering-skills

Treat memory files, summaries, handoff docs, and other agents' reports as unverified claims. Inheritance tells you where to look; only the current checkout tells you what's true. Apply whenever acting on knowledge you did not derive this session.

2026-07-050

name	agent-prompt-doctrine
description	Eight falsifiable rules for prompting CLI coding agents and review lanes, each derived from an observed failing-form vs working-form pair. Use when composing any prompt for a coding agent or reviewer fanout.

Agent Prompt Doctrine

The eight rules

One prompt = one verifiable diff. Whole-package verdict requests get defensive rejections; "review ONLY these 3 files" / "make EXACTLY these changes" completes cleanly.
Bound the evidence set. Inline the exact files/spec text the agent needs; lanes forced to re-read a whole repo burn tokens and return mush.
Embed the acceptance command. Name the exact test command (with interpreter path) that must pass. "It should work" prompts end in the future tense.
Enumerate prohibitions. Agents self-audit against explicit "do NOT reformat / do NOT add error handling / do NOT touch other files" lists. Unstated constraints get violated.
Scope beats format-nagging. Scolding "give a terminal verdict, not one word" half-works; narrowing scope to one diff with evidence inlined actually fixes outcomes.
No status protocols in one-shots. A "send WORKING before long passes" boilerplate caused lanes to END on the heartbeat — "WORKING: ..." as the final answer.
Never respawn an unchanged fanout. Same prompt + same scope = same failure. Change the decomposition, not the dice.
Pre-pay investigation. Implementer prompts contain conclusions (paths, line refs, decided approach), not open questions. Investigation is a separate, cheaper, read-only pass.

Prompt skeleton

<ONE commit-sized task statement>
Files: <exact paths, insertion points, line refs if known>
Spec: <verbatim spec text — do not paraphrase your plan>
Do NOT: <reformat / add defensive code / touch other files / run long test suites>
Must pass: <exact test command>

Timeouts

Full test suites hang agents (observed: a unit run stalled at 17% for 7+ minutes). Agents run focused test files only; the orchestrator runs the broad suite at the review gate.