Dako's AI Builder Notes
⌘K
中EN
System
Color theme
Changsha HubOffline
大可 Dako
大可 Dako大可 DakoAI Builder · Agent PractitionerTurning AI into working software. Practical notes on agent engineering, vibe coding, and commercial delivery.
Platform links coming soon
Projectharnessthe layer yourproject addsRead righttrusted context, it doesn't guessBlockenforceable constraints, stop therecurring mistakesReconnectrecoverable flow, can resume aftera breakPut common sense into the project so AI doesn't run on guessing

The Three Jobs of a Project Harness: Read Right, Block, and Reconnect

Turn the frequent mistakes into constraints instead of scolding AI. A project harness only does three things: read right, block, and reconnect.

The same requirement, one person with well-prepared project context and another with a bare project, even on the same AI, produce wildly different results. The gap isn't in the model; it's in "does the project set a foundation for AI"—that's the project harness.

After reading, you should be able to answer:

  • What does a project harness actually do?
  • What do "read right," "block," and "reconnect" each mean?
  • Why can't "write a harsher prompt" be your way to block AI?

What a project harness is

A project harness is, bluntly, the configuration and tooling in your project that help AI "read right, block, and reconnect": rule files, directory structure, scripts, checks, and how they're organized. It decides how well AI works in your project, not limited by "the model itself."

The three jobs, unpacked

1. Read right (trusted context) Let AI get reliable information about your project instead of guessing.

  • AGENTS rule files: directory, structure, conventions
  • Relevant docs / README / code examples
  • Avoid giving wrong: tell it where source-of-truth is and what to trust

2. Block (enforceable constraints) Turn "the problems that keep happening" into constraints AI can't get around, instead of relying on it to remember.

  • What a machine can judge, make into a script / check / pre-action gate
  • The criterion is "can it continue after violating": can continue = reminder; can't = constraint
  • High-risk actions (push to protected branch, delete DB, touch prod) blocked by a machine

3. Reconnect (recoverable flow) If a brand-new AI (or a new session) takes over mid-way, can it continue using only what's left in the project?

  • Write stage artifacts into files / docs, not just into the chat
  • Handoff contract: input comes from the previous step's artifact, output lands in a file
  • "The swap-the-agent test": assume a new person, only project state remains—can it continue?

Why a "harsher prompt" isn't enough

Many believe that "writing a harsher prompt" keeps AI from messing up, but a prompt is a reminder, not a constraint. To block, wire the machine-judgeable requirements into the execution chain (scripts / checks / gates), not depend on AI "remembering."

  • Reminder: you can continue without it
  • Constraint: you can't move without it

Your harness wants constraints, not declarations.

How the three jobs land

  • Grow from real failures: fix a one-off problem first; when it repeats, turn it into a rule or script anthropic-sdlc
  • Don't pile on a complex harness at the start: start from a minimal baseline, and let every addition have a clear "the failure that triggered it"
  • Automate what a machine can judge: scripts, checks, CI gates are stronger than a doc entry

With the harness in place, AI works in a territory where it "reads right, is blocked properly, and can reconnect." Next is the actual building—Build in Small Steps.

References

  1. The AI-Native SDLC Playbook (Anthropic)