From Prompt to Output: Day-to-Day AI Agent Workflow

infograph

1. Mindset

When the result/output isn’t satisfactory, check the prompt first — unclear instructions are usually the human’s error, not the agent’s
Tell the story clearly, one chapter at a time.

Update ~/.openclaw/SOUL.md and ~/.hermes/SOUL.md if there isn’t, to make the agent more specific
Understand markdown
Linux is ideal; macOS is fine. Avoid running agents on Windows:
- Administrator privilege is too broad — risk of touching the whole OS
- PowerShell is optional for the OS but often required by agents, which adds friction
Learn basic Unix/Linux as a non-root user
Python: basic understanding is nice to have
Result filename convention (OpenClaw and Hermes): YYMMDD + time + topic + subtopic, e.g. 2605211430-report-summary.md
- Avoid spaces in filenames
- Know where to save outputs. Default agent paths: ~/.openclaw, ~/.hermes (config, skills, state, including conversation history and output files)

Write my own most of the time
Read others at https://clawhub.ai, https://www.skills.sh and https://hermes-agent.nousresearch.com/docs/skills
Understand exactly what each installed skill does before you rely on it
Install skills and plugins only when needed — performance and security. Skills: AI chooses when to use. Plugins: auto hooks (e.g. email) without AI choosing.
Be skeptical of long skill files; harder to audit and more likely to hide unwanted behavior

Prefer English for prompts (vs Chinese or others) when you need precision
Use precise wording; cross-check with tools like Gemini or Perplexity if needed
Keep prompts short and precise — saves tokens and reduces errors
Break work from large to small (goal → steps → single tasks)
No duplicated messages in the same session
New task → new session — unrelated history gets loaded as prompt context
Do not keep correcting an old prompt; that confuses the agent. Start fresh instead

Architectural thinking: overall → details, step by step (e.g. build a house: plan → materials → foundation → plumbing, wiring, tiling, paint → finish)
Run independent pieces in parallel (e.g. plumbing, wiring, tiling at the same time)
Break work into small pieces when it cannot be done in one shot; start small
Guardrails: state what must not happen (e.g. “Everything in Python. Do not use Java”)

Let AI review your SKILL.md before you rely on it
Enable verbose mode and watch the test run
SKILL.md: standard skill markdown — YAML frontmatter + declarative human-language instructions