Build-In-Public - Ship-End Suggestions

AI-TUNE Agent Conventions

This read-only page proposes source-safe post candidates from the shipped boundary: root agent notes that make "AI tuning" route to the offline AI-TUNE evidence/review loop instead of direct browser-runtime AI edits. It does not publish posts or write social-ledger records.

bip_phase: implementation source skill: ship-end date: 2026-07-02

Source Basis

Loaded convention path note: local source checkout does not include docs/social-ledger-convention.md or bundled channel convention files. Candidates follow the ship-end BIP page contract and mark this missing local convention path explicitly.

Text And Community Candidates

LinkedIn

recommended
Angle: making AI tuning auditable before it can affect runtime behavior.
Risk level: low
Loaded convention path: missing local social convention docs; fallback to ship-end BIP contract.
Claim-safety notes: process claim only; no performance, model-quality, or launch claim.
Publish precheck: verify repo notes are intended public and do not expose private roadmap details.
I added a small convention to the project: when I ask for "AI tuning," agents should run the offline evidence/review loop, not edit runtime AI directly. The loop runs deterministic local scenarios, writes structured logs, produces candidate review notes, and keeps every candidate as "not promoted" until a human-reviewed promotion path exists.

X

recommended
Angle: concise guardrail around AI automation.
Risk level: low
Loaded convention path: missing local social convention docs; fallback to ship-end BIP contract.
Claim-safety notes: no claim that the AI is improved, only that the workflow is safer.
Publish precheck: avoid naming specific candidate values unless the report is public-safe.
Project convention added: "tune the AI" means run the offline AI-TUNE evidence loop, not patch live browser AI. Simulate locally, review structured logs, keep candidates as evidence, and promote nothing without explicit human-reviewed runtime values.

Bluesky

recommended
Angle: small safety boundary for game AI iteration.
Risk level: low
Loaded convention path: missing local social convention docs; fallback to ship-end BIP contract.
Claim-safety notes: keeps source-safe wording and avoids unverifiable outcome claims.
Publish precheck: confirm no internal agent tooling details should remain private.
Useful game-AI workflow rule: AI tuning requests should produce review evidence first. Runtime behavior only changes after a reviewed promotion step. This keeps tuning from becoming an accidental path for opaque or unsafe behavior to ship.

Mastodon

recommended
Angle: deterministic local testing before live behavior changes.
Risk level: low
Loaded convention path: missing local social convention docs; fallback to ship-end BIP contract.
Claim-safety notes: no external benchmarks or public quality claims.
Publish precheck: use content warning if paired with war-game screenshots.
Added an agent note to keep game-AI work boring in the right way: "AI tuning" now routes to local deterministic scenario runs, structured logs, and candidate review. Runtime AI stays authored until a later promotion task explicitly names the values and files.

Reddit

not-now
Angle: workflow feedback post for game-dev process communities.
Risk level: medium
Loaded convention path: missing local social convention docs; fallback to ship-end BIP contract.
Claim-safety notes: would need a concrete question and public artifact link.
Publish precheck: read community rules and avoid promotional framing.
Possible title: "How do you keep offline AI tuning from accidentally becoming runtime behavior?" Body should ask for process critique, not market the project.

Hacker News

rejected
Angle: process note about deterministic game-AI tuning.
Risk level: high
Loaded convention path: missing local social convention docs; fallback to ship-end BIP contract.
Claim-safety notes: not enough standalone technical artifact for HN yet.
Publish precheck: wait for promotion-packet tooling or a deeper technical write-up.
Rejected for this wrap-up. Save for a later post once the fail-closed promotion gate and run manifest exist.

Video Candidates

YouTube Shorts / TikTok

not-now
Angle: 30-second visual on "AI tuning is evidence first."
Risk level: medium
Loaded convention path: missing local video convention docs; fallback to ship-end BIP contract.
Claim-safety notes: needs screen capture of reports, not just code text.
Publish precheck: wait until promotion-packet tooling creates a clearer before/after visual.
Outline: "I do not let AI tuning patch runtime AI directly." Show command run, summary report, candidate review, and the not-promoted boundary.

YouTube Long-Form

rejected
Angle: full AI tuning pipeline walkthrough.
Risk level: high
Loaded convention path: missing local video convention docs; fallback to ship-end BIP contract.
Claim-safety notes: premature until the promotion gate, expanded guardrails, and run manifest are implemented.
Publish precheck: revisit after the pipeline can safely promote an exact reviewed value.
Rejected for this wrap-up. The current work is a convention and process audit, not enough for a complete technical video.