Stop Fixing the Input

The reframe

Lou opened with a question most of the group recognized: why do we spend so much time optimizing prompts when the model is non-deterministic? Even the best prompt produces slop on some runs. And when a new model release reshuffles behavior, all that input-side tuning has to start over.

His answer: stop fixing the input and build a quality floor on the output side instead.

The system starts from gold-standard examples — 20 to 50 pieces of the best content you’d want to produce, your own work or anyone else’s you’d aspire to match. A command reads through these and derives a scoring rubric: not just grammar and structure, but substance — hook quality, angle, perspective, whether the piece earns its claim. The rubric encodes what makes these pieces worth publishing.

The gate runs downstream of any writing skill. It accepts content plus a content-type label, scores the output against the matching rubric from 0 to 1, and returns what failed. It never edits. If the output scores below threshold, the writing skill reruns. The gate gets harder to fool over time — every use and every correction sharpens the discrimination. Multiple content types, multiple rubrics, each with its own 20 to 50 examples. The rubric encodes your taste, not best-practice defaults.

The library architecture

Connected to the quality gate, Lou described the AIA-lib and harness pattern he’s been developing to solve the token-tax problem. As a skills library grows, embedding everything in a global CLAUDE.md costs 10 to 15 thousand tokens on every inference — paid whether or not the project needs those capabilities.

The solution: one central library on GitHub holds all skills, commands, resources, and agents. Nothing loads automatically. Each project folder contains a harness file that declares only the capabilities it needs. The global CLAUDE.md holds one thing: a pointer to where the library lives. Lou’s mental model: a USB hub — one cable in, ten ports on top, plug in only what this task requires. The token cost stays flat regardless of how large the library grows.

Donald’s computer use discovery

Donald opened his segment with a clean statement: he doesn’t need APIs anymore.

GoHighLevel doesn’t expose the endpoint he needed. Normally that would be the end of it. But Codex’s computer-use feature opens its own Chrome session, navigates to the platform, and clicks through menus as a human would. He handed the task to Codex and watched it navigate the deepest menus of GoHighLevel to build a full inventory of his instance — all tags, all URLs, all APIs. It’s clicking around, even the deepest menus. I felt like: now I really have an assistant.

The parallel execution model he’s landed on: while Codex navigates GoHighLevel, Donald runs Claude for the next planning step. Neither model burns idle while the other works. He described it as feeling like a team of three.

Don Back’s negotiation rehearsal

Joanna’s closing question — what has been the most groundbreaking AI innovation you’ve seen? — produced the session’s best story.

Don Back is co-chair of a not-for-profit board facing a governance issue festering for three years. He recorded the last meeting where everyone stated their position, then loaded the transcript, the Act, the bylaws, his objective, his negotiating strategy, and seven proposed actions — only one of which he actually wanted — into Claude in back-and-forth debate mode. He tested every position, audited every argument, tuned the strategy.

Result: he walked into the executive committee meeting and it rolled out exactly as it was role-played. He’ll never tell the other board members how he prepared. It wasn’t telling me the truth. But it was enabling me to think it through and get better and better at honing what I wanted to achieve.

Kasimir extended this into a principle: the default AI mode is search-and-summarize. The higher use is as a thinking partner — but only if you break the sycophancy first. His technique: give AI personas instructions that they will not change their position without data. The data-gate forces you to actually reason rather than assert.

Kasimir’s avatar video pipeline

Kasimir has been building a serialized avatar video production pipeline using HeyGen’s API — three custom characters, twenty scripts forming a story arc, and a workflow that generates scripts, uploads to HeyGen, handles pauses and captions automatically, and downloads the rendered video. The plan: pay for one month, batch-produce all twenty videos, release on a schedule. The AI does the repetitive production; he does the story design and narrative arc.

He also mentioned a quality gate of his own: a virtual board of advisors scores each post, and when it clears the threshold he reviews it manually before publishing.

What to try before next session

Pick one AI interaction you repeat weekly — something you could theoretically automate but haven’t. Collect five to ten examples of that output you’re proud of. Ask Claude: Read these examples. What distinguishes the best ones? Generate a scoring rubric I can use to evaluate future versions.

That rubric is the seed of your quality gate.