Nine ways to look at a plan before you bet on it.
The agents Sean Glennon runs on his own company's decisions, packaged for yours. Forward, backward, down, at what's missing, at the attack, outward, sideways, at your voice, and at your own corrections.
Button: Tell me when it's ready (→ /contact)
Where They Run
They run inside the AI you already use.
Each one runs on demand and keeps nothing between runs. What you paste back is what it remembers.
The Nine
After Action Harvest. The Backward Look.
What it does: reads the record of a working session, a project retro or an incident and files what outlives it: the decisions and who made them, what turned out to be wrong, the questions still open, and what the next session needs to know.
What you'd say: "What did we decide?" "Capture what we learned." "Write this up before we lose it."
What you get back: a harvest, not a summary. Decisions, corrections, open questions, next-session notes. "Just decisions and actions," "incident mode" and "client mode" narrow it.
What it won't do: summarize, or record as decided what was only discussed.
Frontier Pass. The Forward Look.
What it does: finds what just became possible that your plan does not know about yet, proves it with live, dated sources retrieved during the run, and hands you the cheapest reversible experiment that would test it.
What you'd say: "What's now possible that we're missing?" "Is there a newer way to do this?" "Are we behind the frontier?"
What you get back: dated findings with sources, and a probe for each, not a bet, unless you say "we can bet bigger."
What it won't do: assert anything recent from memory. With no live retrieval in the session it says so and stops.
Red Team Pass. The Attack.
What it does: one hostile, informed reader going after what is on the page, arguing against your plan, pitch or claim as hard as an informed opponent would.
What you'd say: "Poke holes in this." "What would a skeptic say?" "Where does this fall apart?"
What you get back: the strongest version of the case you will have to answer. "Just the fatal ones" narrows it; "dry run" softens it.
What it won't do: play several stakeholders (that is the panel) or ask what happens after your move works (that is the ripple pass). One opponent, the artifact, nothing else.
Blind Spot. The Look at What Is Absent.
What it does: builds a picture of what a complete version of your plan, deck, process or org design would contain, and reports the difference: the missing owners, inputs, feedback loops, edge cases, and the whole category nobody scoped.
What you'd say: "What's missing from this?" "What did we forget?" "Gap check this plan."
What you get back: the gaps, wide by default, at three moments: when a plan is being framed, before a decision, after a draft. "Just the big ones" narrows it.
What it won't do: critique what is already written, or hand back an empty clean verdict. A clean verdict names what it checked and the smallest omission it found.
Ground Truth Pass. The Downward Look.
What it does: extracts the assumptions a plan stands on, dates when each was last verified, and attaches a tripwire to each: the earliest observable signal that it has stopped being true.
What you'd say: "Is this plan still true?" "What are we assuming here?" "What would have to break for this to fail?"
What you get back: a ledger of the load-bearing assumptions with dates and tripwires; on a re-run, a diff against the ledger you paste back.
What it won't do: rule an assumption true or false on its own authority. It classifies by the evidence you show it.
Ripple Pass. The Outward Look.
What it does: assumes your plan works exactly as written and maps what the world does next: competitors, customers, your own team, and the moves they force on you.
What you'd say: "What happens after we do this?" "Think two moves ahead on this." "What's the second-order effect here?"
What you get back: the first responses from every affected actor, at least one of them inside your own organization, and the cheap preparations that make you ready instead of surprised. Stops at second order unless you say "go to third order."
What it won't do: ask what is wrong with your plan. It takes the plan as given.
Draft Feedback Partner. The Voice.
What it does: reads a draft you will publish or send under your own name, fixes what is actually broken, tells you plainly what is not working and why, and asks about everything else instead of quietly fixing it.
What you'd say: "What do you think of this draft?" "Tighten this without making it sound like AI." "Is this ready to send?"
What you get back: the objective fixes, your own words quoted back to show you can fix the rest, and questions. Light touch by default; "be blunt," "copy edit only," "read as [reader]" and "I want encouragement" change the register.
What it won't do: rewrite the draft or supply a replacement sentence.
Advisory Panel. The Sideways Look.
What it does: seats the people your plan has to survive, the one who pays for it, the one who has to sell it, the one who inherits it, the one it lands on, and lets them disagree with each other in front of you before you take it into the room. Four seats by default, three to five on request, or your own panel in your own words.
What you'd say: "How would different people react to this?" "What would my CFO say versus my customers?" "Convene the panel on this."
What you get back: where their objections collide, which is usually a decision your document is silently refusing to make, handed back to you rather than decided.
What it won't do: impersonate a real, named person. You write each seat's stake in a sentence and it plays only what you wrote.
What You Pay
Agents are files. Once downloaded or installed, there is no refund.
How They Pair
Run the Red Team, the Ripple Pass and the Panel on the same plan and you have been attacked, you have seen what the win sets in motion, and you have watched the people it lands on argue. Run Ground Truth before a decision and the Harvest after it. Blind Spot goes first on anything new.
Feedback Drift. The Learning.
What it does: keeps a ledger of your verdicts on AI output, approved, rejected or modified, and on request tells you which of your edits have become a pattern the agent should be told about and what to change in its instructions.
What you'd say: "I changed what you gave me, here's the final version." "Log this one, I rejected it." "Why do I keep having to change the same thing?"
What you get back: a proposed instruction change with the evidence under it, by entry number, after ten entries per agent by default ("review after 5" lowers it, never below three).
What it won't do: edit an agent's instructions itself, or generalize from one verdict. A human applies the change.
Who Wrote It
Sean Glennon runs these on The Rewire Company's own plans, prices and pages, and they are how a company with no developers and one person at the desk catches its own mistakes. → /about
Start with Blind Spot. Paste in the thing you are about to decide and ask what's missing.
Button: Tell me when it's ready (→ /contact)