I built a playbook inside Founders OS that hunts for consulting contracts. It searches listings, scores each one against what my two-person shop actually wins, writes tasks for the good ones, and drafts proposals for the exceptional ones. Standard automation so far.
The part that matters is the last step of every run: a retrospective. After I act on a batch, the playbook asks me what the scores got wrong. What did I pass on that scored high? What did I chase that went nowhere? What red flags or stats do the no response bids have in common. Then it writes those lessons back into its own scoring rubric. The next run starts smarter than the last one.
Here is what that loop actually taught it, in order, from real runs.
Run one, it fired a strong proposal at a $25K build with fifty-plus proposals. Never got viewed. Client hired someone else. Same thing happened on the next crowded listing. The retro turned that into a rule: a crowded slate where the client is already deep in their funnel is not competitive, it is dead. Stop spending effort there.
A few runs later it was still docking good clients just for having thin payment history. I told it that was wrong, thin history is fine if payment is verified. It corrected the rubric. But then it learned the sharper version on its own runs: do not look at whether a client has history, look at what they paid past hires. A client whose prior hires were all sub-ten-dollars-an-hour, or whose past jobs were all credit-repair gigs, is a red flag no matter how good the listing reads.
Then it learned to prefer fixed-price over hourly, because that is how we actually want to get paid. It learned that "already interviewing a dozen candidates" beats rate and fit every time, skip it. It even fired two search terms that kept returning the wrong domain entirely.
None of that came from me sitting down to write rules. It came from the system reading its own outcomes and adjusting.
The result is not a flood of wins. It is a much tighter shortlist and far fewer wasted bids, and it has started landing wins on and more proposal engagement as it adapts. For a two-person shop where my time is the whole budget, not chasing the dead listings is the win.
One more thing that mattered to me. All of that learning, every lesson about what converts and what does not, lives in my own database. Founders OS is open source and self-hosted, runs over stdio against my own Postgres. The memory of what works for my business is not sitting in someone else's SaaS that I would lose if I left.
GitHub: https://github.com/OurThinkTank/founders-os
Site: https://foundersmcp.com
npm: @ourthinktank/founders-os
If you run a small shop, how do you decide what work to chase? I am curious whether anyone else has tried to make that decision learn from its own misses.
Top comments (0)