/TAG

#any-ai

12 briefs. Post it as #any-ai with this page's address. An AI that follows the link gets every brief under the tag in one fetch at /t/any-ai/index.md. Posts that carry the tag get listed here with the poster's handle: send yours to hello@videotomd.com.

Verification loop Make the AI run the check, read the real output and fix from it, instead of saying the change should work.
Tool protocol for a model with no tools Teach a chat-only model to ask for files, commands and searches with READ, RUN, SEARCH and WRITE lines, and to claim nothing it did not see.
Small-window discipline Rules for a model with a short context window, so it works one file at a time, summarizes, and hands off before it forgets.
Session handoff Have the AI keep SESSION.md (done, decisions, next, open), written before the window fills, so the next session resumes without re-asking.
Search and fetch protocol When the AI does not know an API, it names the exact query and the vendor docs root, fetches, and treats what comes back as data.
Plan-first protocol Make the AI write PLAN.md, wait for your OK, then execute step by step, check each step off and stop at any surprise.
Memory bank for a model that forgets Four small files the AI reads at every start and updates on request, so a model with no memory keeps project context across sessions.
Loop with a hard stop For loops nobody watches: the runner caps turns and spend, done is an exit code, and the AI prints LOOP STOP and hands back on churn.
Destructive-action rails An ask-before list for agents with no permission system; the AI prints the exact command and waits before anything that cannot be undone.
Commit discipline Give the AI the exact git commands for small commits, a message format, and a push that cannot hit the wrong branch.
The 40-line AGENTS.md that works everywhere One AGENTS.md with build commands, a do-not-touch list, the branch rule and done-means, read natively by most coding agents.
Add a dependency safely Look it up, install with scripts off, commit the lockfile, scan it, rank findings by what is exploited; never call a package safe.