Why One‑Shot AI Prompts Fail and How 19 Iron Rules Build a Factory‑Style Workflow

Why single‑turn AI prompts fall short

When users ask an AI to write a 3000‑word article in one conversation, the output often feels generic, repetitive, or requires endless revisions. The core issue is not AI capability but the "one AI + one conversation" model, which cannot manage complex, multi‑step tasks.

Harness: an AI "factory" system

Harness, deployed inside ByteDance’s AI programming tool, treats AI agents as a coordinated team rather than isolated bots. It defines six distinct roles:

Project Manager : schedules the right person at the right time, without writing or coding.

Material Specialist : extracts key points, quotes, data, and cases from raw inputs.

Planning Manager : creates a writing strategy—angle, narrative thread, and paragraph pacing.

Lead Writer : turns the plan and material into a first draft.

Editor : performs a four‑dimensional review (originality, impact, factuality, style).

Chief Editor : makes the final decision—publish, revise, or rewrite.

These roles are arranged in a linear stage sequence (S0‑S6) that mimics a production line, ensuring each output passes to the next stage without back‑tracking.

Five‑step factory setup

Measure : The system asks for the user’s domain, desired output, and quality standards before proceeding.

Show sample : It generates a workflow diagram; the user can add or remove roles.

Define standards : At least eight quality rules (e.g., max paragraph length, citation requirement) and six automatic checks are created.

One‑shot generation : After confirmation, the system produces all configuration files, role definitions, and scripts in a single run.

Self‑inspection : Automatic scripts verify every rule; each answer is a binary "yes" or "no".

Key iron rules (selected examples)

"Do not start before the requirement is clear."

"Do not use vague completion language; status must be either generated, skipped, or failed."

"Do not skip self‑inspection; every output must pass the automated checks."

"Do not modify the inspection scripts; they are locked in the scripts directory."

"If a stage is rejected three consecutive times, the pipeline pauses for human intervention."

These constraints prevent the endless back‑and‑forth that users experience with ordinary chat‑based AI.

Token management

Because each AI has a limited context window measured in tokens, Harness forces the Project Manager to produce a concise context summary (≤500 Chinese characters) for each hand‑off, containing only the current step, a directory of produced artifacts, and the next step’s requirements. This reduces token usage from thousands to a few hundred while keeping full files on disk for on‑demand lookup.

Two domain examples

Content creation : The workflow becomes "topic planning → material collection → drafting → editing → final review → formatting" with ten quality rules such as "unique titles" and "one key quote per 500 words".

Software development : The workflow expands to eight stages—"requirements → architecture → gate review → implementation → code review → testing → performance evaluation → deployment"—and includes rules like "no bare catch" and "test coverage must not drop".

Both share the same skeleton: a goal definition stage, a delivery stage, and domain‑specific execution and review stages, with automatic checks that can roll back upstream work when problems are detected.

Self‑evolving factory

An additional "Evolution Agent" monitors the pipeline, detects bottlenecks (e.g., repeated rejections), and proposes incremental improvements. It only changes a few critical steps at a time and requires human approval before applying changes, ensuring controlled evolution.

Conclusion

Harness does not make AI smarter; it provides a disciplined method to organize multiple AIs. Its 19 iron rules, staged pipeline, token‑efficient hand‑offs, and self‑evolving agent make complex tasks reproducible, evolvable, and trustworthy.