Execute a PRIVATE_BATTLE.md

PRIVATE_BATTLE.md is a file-first way to define a battle spec alongside your code — commit it to your repo and reproduce the same AI competition anywhere.

Time: ~10 minutes
Prerequisites: lf CLI installed, provider API keys set (e.g. ANTHROPIC_API_KEY, OPENAI_API_KEY)

Step 1 — Create a PRIVATE_BATTLE.md

Create the file in your project directory:

yaml

---
kind: private_battle
name: "TypeScript Haiku Duel"
slug: ts-haiku-duel
id: ts-haiku-duel-001
participants:
  - type: model
    ref: claude
    provider: anthropic
    model: claude-sonnet-4-6
  - type: model
    ref: gpt4o
    provider: openai
    model: gpt-4o
    key_var: OPENAI_API_KEY
---

## Purpose

Compare how two frontier models write creative TypeScript-themed haiku.

## Participants

| Slot | Model | Provider |
|---|---|---|
| A | claude-sonnet-4-6 | Anthropic |
| B | gpt-4o | OpenAI |

## Evaluation

Judged by brevity, syllable count, and technical accuracy. AI judge determines the winner automatically after execution.

## Report

Results saved to `ts-haiku-duel.result.md` after execution.

Required sections: Purpose, Participants, Evaluation, Report — the validator enforces these.

Step 2 — Validate (no AI called)

bash

lf battle run PRIVATE_BATTLE.md

Expected output:

✔ Simulated private battle TypeScript Haiku Duel
Participants: 2
JSON report: <user-runtime>/workspaces/<id>/runs/ts-haiku-duel.json
Markdown report: <user-runtime>/workspaces/<id>/reports/ts-haiku-duel.md

Tip: participants have provider/model set — run with --execute to actually call AI:
  lf battle run PRIVATE_BATTLE.md --execute

No API calls. Use this to check syntax and required sections before burning tokens.

Step 3 — Execute

bash

lf battle run PRIVATE_BATTLE.md --execute

Both contenders stream simultaneously to your terminal:

[A] Types guard each boundary —
[B] Semicolons fade at last;
[A] async waits in silence.
[B] generics hold the world tight.

When both finish, the AI judge evaluates automatically:

✔ Execution complete in 5891ms.
Results: ts-haiku-duel.result.md
JSON:    ts-haiku-duel.result.json

AI judge evaluating… (~600 input + ~100 output tokens via ANTHROPIC_API_KEY)
✔ Winner: Contender A
  Rationale: Better 5-7-5 syllable structure with stronger technical metaphor.
  Judge: anthropic/claude-haiku-4-5 (42 tokens)

Override: lf battle local vote a1b2c3d --slot A|B|draw

The AI judge verdict is written to ts-haiku-duel.result.md under ## Judge Verdict.

Step 4 — Read the results

bash

cat ts-haiku-duel.result.md

The file contains both outputs under ## Contender A and ## Contender B headers, plus a ## Judge Verdict section with the winner and rationale.

The .result.json contains raw output, token counts, timing, and the battle ID.

Step 5 — Override the verdict (optional)

The AI judge verdict stands unless you disagree. To override:

bash

lf battle local vote --slot B --rationale "Preferred the imagery in B"

Human votes are always authoritative — they override the AI judge in lf battle local status.

To skip the AI judge entirely:

bash

lf battle run PRIVATE_BATTLE.md --execute --no-judge

Step 6 — Optionally push to cloud

To share results publicly:

bash

lf battle local push --slug "ts-haiku-duel-cloud"

This creates a cloud draft with the same title and task. Outputs are not auto-uploaded; use the web UI to add them as contender submissions.

Using `key_var` for per-participant key override

If you store different API keys under non-standard env var names, set key_var in the participant:

yaml

participants:
  - type: model
    ref: my-claude
    provider: anthropic
    model: claude-sonnet-4-6
    key_var: MY_PROJECT_ANTHROPIC_KEY   # reads this var instead of ANTHROPIC_API_KEY

What to try next

More than 2 contenders — Only the first two participants with provider+model run; add more for rotation testing
Ollama — Set provider: ollama, model: llama3.2, no key_var needed
Cloud battle from same spec — lf battle local push --slug <slug> after running

Execute a PRIVATE_BATTLE.md ​

Step 1 — Create a PRIVATE_BATTLE.md ​

Step 2 — Validate (no AI called) ​

Step 3 — Execute ​

Step 4 — Read the results ​

Step 5 — Override the verdict (optional) ​

Step 6 — Optionally push to cloud ​

Using key_var for per-participant key override ​

What to try next ​

See also ​

Execute a PRIVATE_BATTLE.md

Step 1 — Create a PRIVATE_BATTLE.md

Step 2 — Validate (no AI called)

Step 3 — Execute

Step 4 — Read the results

Step 5 — Override the verdict (optional)

Step 6 — Optionally push to cloud

Using `key_var` for per-participant key override

What to try next

See also