Harness Learn

How it works

You invoke the launcher:  /implement "Add dark mode"
                              ↓
Launcher creates a state file and hands off to the engine
                              ↓
Engine loops through phase skills:
  [phase 1] → [phase 2] → ... → [phase N] → COMPLETE
                              ↓
Each phase is a focused 40-140 line skill that:
  - Knows exactly what to do (real commands, not placeholders)
  - Has a checklist the engine validates before advancing
  - Records progress to a conversation file (survives session drops)

Why phases matter

A single big prompt ("implement this feature, test it, create a PR") loses agent attention by the end. By phase 7, the agent has forgotten phase 2's instructions. Phase skills keep each turn focused — the agent sees one phase at a time.

The improvement loop

The harness gets better with use:

/harness-setup (discover your workflow, generate skills)
      ↓
Real work (/implement on actual tasks)
      ↓
/harness-retro (review what happened, fix skill gaps)
      ↓
Back to /harness-setup or next round of work

Pause here. Ask:

"That's the general concept. Want me to show you how it works specifically on this project?"

Layer 2: This Project's Harness

Read the project's harness files and present them in order.

Step 1: Read the harness

Read these files (skip any that don't exist):

1. HARNESS.md — overview and architecture
2. .harness/lifecycle.md — the discovered workflow
3. CLAUDE.md (or .cursorrules, GEMINI.md) — the Harness Context section
4. List all phase skills in the project's skill directory (harness-* and the launcher)

Step 2: Present the workflow

"Here's how this project ships software:"

Show the lifecycle as a table:

Then show the profiles:

"Not every task goes through all phases:"

Step 3: Walk through each phase

For each agent-owned phase, read the phase skill and explain:

[Phase name] — [one sentence on what it does]

• Produces: [artifact]
• Verified by: [method]
• Key commands: [actual commands from the skill]
• Gate: [what must be true to proceed]
• Checklist: [items the engine validates]

For human-owned gates, explain:

[Phase name] — This is where YOU step in.

• The engine stops and waits for your approval.
• You [what the human does — review PR, test UI, approve deploy].
• When you approve, the engine continues to the next phase.

Step 4: Show a real example (if available)

Check .harness/conversations/ for past implementation records. If any exist, pick the most recent completed one and narrate:

"Here's what happened the last time someone used this harness:"

• Task: [from conversation file]
• Profile: [profile used]
• Phases completed: [list with durations if available]
• What was produced: [commits, PR, evidence]
• Any friction: [issues from the conversation file]

If .harness/retros/ has retro records, mention:

"The harness has been through [N] retro rounds. Recent improvements: [list changes from retros]"

If no conversation files exist:

"This harness hasn't been used yet — no past examples to show. Want to do a dry run?"

Layer 3: Hands-On Walkthrough

Ask:

"Want to try a dry run? Pick a task (or I'll suggest one) and I'll walk you through what each phase WOULD do — without actually running it."

If they say yes:

1. Ask for a task description (or suggest a small one from the backlog if PM tool is available)
2. Determine which profile would apply
3. For each phase in that profile, explain what the agent WOULD do:
   • "In the [phase] phase, the agent would [action]. It would run [command] and check that [gate condition]."
   • Show what the state file would look like at each transition
4. Point out where human gates would pause the workflow
5. At the end: "That's the full lifecycle. Ready to try it for real? Run /implement [task]"

If they say no:

"No problem. When you're ready, run /implement with a task description to start. The engine handles the rest."

Adapt to the Audience

• Developer new to the repo: Focus on the commands and where their code goes in the workflow. Skip philosophy.
• PM or non-technical: Focus on the phases, gates, and what they'll see (PRs, status updates). Skip commands.
• Agent joining a team: Focus on the state file, conversation recording, and escalation rules. They need to know when to stop and ask.
• Someone who's used harnesses before: Skip Layer 1, go straight to Layer 2.

Ask at the start:

"Before we start — are you a developer, PM, or agent? And have you seen a harness workflow before?"

This shapes how deep each layer goes.