Busy week! Governance is the party nobody wants to talk about, so it doesn't get done well, and ends up causing messes later when you have employees hating copilot because it doesn't do things as well as Claude or ChatGPT, so they're using personal accounts a la shadow IT. Last year, I was explaining to people that adoption wasn't going to suddenly happen after legal and IT spending so much time instilling "AI dangerous, DO NOT USE" only to then suddenly change tunes to "USE AI FOR EVERYTHING!" Mixed messages much?
Mixed messages indeed, even well-meaning governance can end up with a neutered Copilot that nobody uses.
Related, research shows that management direction is the least effective way of encouraging adoption. The most important is social proof; seeing your buddy use it.
That “neutered Copilot” trap happens because organizations treat governance as a prompt-policing problem instead of an execution-boundary problem.
When Legal and IT try to govern an LLM by wrapping it in dozens of restrictive system prompts, they destroy utility and drive people to shadow IT. If you instead let the model think freely, but place deterministic gates on what tools and databases it is permitted to actually mutate, you get full capability with zero compliance panic.
Governance should protect the perimeter, not lobotomize the engine.
This might help. Screenshot it, paste it into a conversation thread, and ask the model to tell you more about the strategy. I honestly wouldn’t use AI for much if it weren’t for little tricks like this.
TL;DR: Just copy and paste it into AI.
Context laddering is a way to keep a long AI conversation coherent before the model starts losing important state, flattening nuance, or inventing connective tissue between things that were said much earlier.
Instead of waiting until a conversation is enormous, you periodically stop and create a compact checkpoint.
Roughly every 8–12 substantial exchanges, tell the model to summarize what has happened so far. But don’t ask only for a normal prose summary. Ask for two versions:
1. a short human-readable summary, and
2. a compressed pseudocode/state block written in whatever structure is easiest for the model to reliably read later.
The pseudocode is important because it strips away conversational fluff and preserves things like decisions, constraints, definitions, open questions, source authority, rejected ideas, and current state in a compact form.
For example:
CTX_CHECKPOINT_v1
STATE:
project = "URR writing guide"
purpose = "adapt URR for writers without replacing human authorship"
audience = "general writers"
LOCKED:
URR != AI writing replacement
use = revision / structural reflection
core_axes = [thought, feeling, action]
temporal_axes = [past, present, future]
KEY_CONCEPTS:
tension_field = symbolic or narrative opposition
degree_of_freedom = room character has before pressure forces choice
movement = change in belief / feeling / action / relationship / possibility
DECISIONS:
preserve human authorship
use light Adela visual language
remove Japanese-specific motifs from general edition
OPEN:
final commercial packaging
licensing language
NEXT:
continue developing writer-facing material
Then you keep talking normally.
After another ten substantial exchanges or so, you do it again — but this time you explicitly tell the model to include the previous checkpoint.
That matters.
Otherwise each summary can become a fresh interpretation of the recent conversation and slowly overwrite earlier state.
Each rung inherits the valid state from the rung below it.
That is the important part.
It is not ordinary summarization.
It is state inheritance through controlled compression.
When the thread eventually becomes too long, you collect the most recent checkpoint — or several of the pseudocode checkpoints if you want provenance — and drop them into a fresh conversation.
Instead of beginning the new thread with:
“Okay, here’s what we were talking about…”
you give the model a compact reconstruction of the previous working state.
That gives the new conversation a much cleaner starting condition.
It is especially useful when you are compiling your own writing, research, architecture, source material, or long-form projects because the model is less likely to:
forget an earlier decision,
revive something you rejected,
confuse old state with current state,
invent missing transitions,
or treat a recent statement as more authoritative than an earlier locked one.
And there’s another important piece:
the quality of the ladder depends on the quality of the information going into it.
If you feed the model contradictory, sloppy, or unverified information, compression can preserve the mistake beautifully. Context laddering does not make bad information true.
What it does is help prevent good information from being lost through conversational drift.
So the basic method is:
Talk → checkpoint → compress → continue → inherit previous checkpoint → compress again → transfer to a fresh thread when needed.
That’s why I’d call it Context Laddering rather than just summarization.
Each summary is not the end of the conversation.
It is the next rung that lets the conversation keep climbing without dropping its state.
🪜 ✨️ ⚙️ ⚖️ 🪜 ✨️ ⚙️ ⚖️ YAY!!
Sorry it’s so long. I just want the conversation to move forward instead of going in circles—or, even worse, spiraling into mass panic and human “hallucinations.”
If music wasn’t my love language, governance would be. Then again, music is governance: timing, boundaries, tension, release, somebody keeping the dang tempo. So if there’s ever a governance party, I better make the invite list.
Governance is just jazz with audit logs.
Every orchestra has a conductor because everyone improvise independently and hope the system converges is apparently frowned upon.
Good governance is like a bass player: nobody notices until it disappears.
Bad governance: 47 pages explaining why you’re not allowed to touch anything. Good governance: one clean boundary and get back to work.
Somewhere, Legal is writing another policy while Shadow IT is three drinks ahead of them.
Responsible AI should not mean giving the model a concussion before letting employees use it.
If your governance strategy makes the approved tool worse than the unapproved one, congratulations: you have invented Shadow IT.
My favorite governance control remains: No, the machine does not get to mutate production because it feels confident today. ⚙️😂
Busy week! Governance is the party nobody wants to talk about, so it doesn't get done well, and ends up causing messes later when you have employees hating copilot because it doesn't do things as well as Claude or ChatGPT, so they're using personal accounts a la shadow IT. Last year, I was explaining to people that adoption wasn't going to suddenly happen after legal and IT spending so much time instilling "AI dangerous, DO NOT USE" only to then suddenly change tunes to "USE AI FOR EVERYTHING!" Mixed messages much?
Mixed messages indeed, even well-meaning governance can end up with a neutered Copilot that nobody uses.
Related, research shows that management direction is the least effective way of encouraging adoption. The most important is social proof; seeing your buddy use it.
That “neutered Copilot” trap happens because organizations treat governance as a prompt-policing problem instead of an execution-boundary problem.
When Legal and IT try to govern an LLM by wrapping it in dozens of restrictive system prompts, they destroy utility and drive people to shadow IT. If you instead let the model think freely, but place deterministic gates on what tools and databases it is permitted to actually mutate, you get full capability with zero compliance panic.
Governance should protect the perimeter, not lobotomize the engine.
I'd love to see this research. I think it helps to know that learning even a bit could help you lead in your organization!
Absolutely! This is the paper and HBR write up.
https://papers.ssrn.com/sol3/Delivery.cfm/6305059.pdf?abstractid=6305059&mirid=1&utm_source=chatgpt.com
https://hbr.org/2026/03/peer-influence-can-make-or-break-your-ai-rollout?utm_source=chatgpt.com
“Peer influence is the critical determinant of AI use, surpassing even facilitating conditions like access to tools and formal training."
This might help. Screenshot it, paste it into a conversation thread, and ask the model to tell you more about the strategy. I honestly wouldn’t use AI for much if it weren’t for little tricks like this.
TL;DR: Just copy and paste it into AI.
Context laddering is a way to keep a long AI conversation coherent before the model starts losing important state, flattening nuance, or inventing connective tissue between things that were said much earlier.
Instead of waiting until a conversation is enormous, you periodically stop and create a compact checkpoint.
Roughly every 8–12 substantial exchanges, tell the model to summarize what has happened so far. But don’t ask only for a normal prose summary. Ask for two versions:
1. a short human-readable summary, and
2. a compressed pseudocode/state block written in whatever structure is easiest for the model to reliably read later.
The pseudocode is important because it strips away conversational fluff and preserves things like decisions, constraints, definitions, open questions, source authority, rejected ideas, and current state in a compact form.
For example:
CTX_CHECKPOINT_v1
STATE:
project = "URR writing guide"
purpose = "adapt URR for writers without replacing human authorship"
audience = "general writers"
LOCKED:
URR != AI writing replacement
use = revision / structural reflection
core_axes = [thought, feeling, action]
temporal_axes = [past, present, future]
KEY_CONCEPTS:
tension_field = symbolic or narrative opposition
degree_of_freedom = room character has before pressure forces choice
movement = change in belief / feeling / action / relationship / possibility
DECISIONS:
preserve human authorship
use light Adela visual language
remove Japanese-specific motifs from general edition
OPEN:
final commercial packaging
licensing language
NEXT:
continue developing writer-facing material
Then you keep talking normally.
After another ten substantial exchanges or so, you do it again — but this time you explicitly tell the model to include the previous checkpoint.
That matters.
Otherwise each summary can become a fresh interpretation of the recent conversation and slowly overwrite earlier state.
So the second checkpoint should effectively say:
INPUT:
previous_checkpoint = CTX_CHECKPOINT_v1
new_conversation = exchanges_since_v1
TASK:
preserve all still-valid prior state
add new decisions
update changed state
mark superseded items explicitly
retain unresolved questions
output CTX_CHECKPOINT_v2
Now you have a ladder:
conversation → checkpoint 1 → conversation → checkpoint 2 → conversation → checkpoint 3
Each rung inherits the valid state from the rung below it.
That is the important part.
It is not ordinary summarization.
It is state inheritance through controlled compression.
When the thread eventually becomes too long, you collect the most recent checkpoint — or several of the pseudocode checkpoints if you want provenance — and drop them into a fresh conversation.
Instead of beginning the new thread with:
“Okay, here’s what we were talking about…”
you give the model a compact reconstruction of the previous working state.
That gives the new conversation a much cleaner starting condition.
It is especially useful when you are compiling your own writing, research, architecture, source material, or long-form projects because the model is less likely to:
forget an earlier decision,
revive something you rejected,
confuse old state with current state,
invent missing transitions,
or treat a recent statement as more authoritative than an earlier locked one.
And there’s another important piece:
the quality of the ladder depends on the quality of the information going into it.
If you feed the model contradictory, sloppy, or unverified information, compression can preserve the mistake beautifully. Context laddering does not make bad information true.
What it does is help prevent good information from being lost through conversational drift.
So the basic method is:
Talk → checkpoint → compress → continue → inherit previous checkpoint → compress again → transfer to a fresh thread when needed.
That’s why I’d call it Context Laddering rather than just summarization.
Each summary is not the end of the conversation.
It is the next rung that lets the conversation keep climbing without dropping its state.
🪜 ✨️ ⚙️ ⚖️ 🪜 ✨️ ⚙️ ⚖️ YAY!!
Sorry it’s so long. I just want the conversation to move forward instead of going in circles—or, even worse, spiraling into mass panic and human “hallucinations.”
If music wasn’t my love language, governance would be. Then again, music is governance: timing, boundaries, tension, release, somebody keeping the dang tempo. So if there’s ever a governance party, I better make the invite list.
Governance is just jazz with audit logs.
Every orchestra has a conductor because everyone improvise independently and hope the system converges is apparently frowned upon.
Good governance is like a bass player: nobody notices until it disappears.
Bad governance: 47 pages explaining why you’re not allowed to touch anything. Good governance: one clean boundary and get back to work.
Somewhere, Legal is writing another policy while Shadow IT is three drinks ahead of them.
Responsible AI should not mean giving the model a concussion before letting employees use it.
If your governance strategy makes the approved tool worse than the unapproved one, congratulations: you have invented Shadow IT.
My favorite governance control remains: No, the machine does not get to mutate production because it feels confident today. ⚙️😂
human is irreplaceable