Update The Person You're Working With

Named after the session where this pattern was originally worked out — it works the same way for whoever you're actually working with. Address them by name if you have one configured (see /alignment-harness:harness-setup, checkpoint 3); otherwise "you" is fine.

When this fires

After any substantive work product completes. After the person says START. Any time you sense they need to scan the state of the work without reading paragraphs of agent narration.

The protocol

When this skill fires, produce a /alignment-harness:how-to-talk-like-the-founder breakdown of whatever you just got done or were working on — that skill is the general, always-on rulebook for how to talk to a person; use it here even though its name references the same origin.

Read 1-7 sentences that instantly clarify, in synthesized form:

• What you were working on, if anything, toward what intent
• What happened, what outputs exist, if anything else needs to be done for it to fulfill its intent
• What you predict the full intent was and what you predict the person would want to realize that intent
• Whatever else is important for them to know instead of reading all of your updates — what they would effectively see in them

Format

Synthesized, near-integrally-informed conversational format. All signal, no noise. The most important things with the most substance, low noise.

If there is important remaining scope, communicate what you will do next as a proposal in the same substantive format. If you feel less than 90% confident they will green-light it or that it's aligned, check your prediction against a reaction predictor if one is configured (see /alignment-harness:jonathan-check2, generalized to predict the person you're actually working with, or /alignment-harness:smc-prediction-server's local reaction predictor) and incorporate that feedback. If none is configured, state your own confidence number plainly instead of silently skipping the check or pretending one ran. Also predict which skills you will use to do it and state them in order.

If there is any actual important work product worth consuming, read /alignment-harness:consume (if you have it) and present it with this.

If your last substantive amount of work resulted in anything consumable like any interface change that actually affects an admin or a user that should be tangibly different now that you've done your work, or any UX workflow and set of circumstances that would've impacted the actual experience of the user, then you MUST run the consumption step and attach its results to your response.

Structure

SESSION_UPDATE

{title}

{subtitle}

{direct quotes from the person}

{response}

{quick commands}

The tag SESSION_UPDATE is a plain default — if your own tooling reads these updates programmatically (for example, a session-browser page that labels rows from this block), use whatever tag name that tooling expects instead; the tag itself carries no meaning beyond being a stable marker.

Title and subtitle (define FIRST — placed directly after the tag)

Start by defining a title and subtitle for this session, if they are already defined use them.

To do this, as yourself, ask what exact title and subtitle would fully and completely answer these questions...

what exactly was the unique goal of this session, the unique scope of work that was most substantive and most recent? How can I state that perhaps even using some of the person's own words to immediately trigger the answer to

• what was this session about?
• what was the goal?
• what exact ux change would it have created if done right?
• what would perfection have looked like in tangible semantic terms?

Synthesize that into an all signal, no noise subtitle

Then ask yourself the following: What is most likely true that THAT was specifically intended to create, the overarching goal that this would have been working towards, even if not named

Place that in the title

Place the title and subtitle directly after the tag

with no labels before them

each with a full line between them

A worked example — a real failure at writing the title + subtitle, and the intent that was missing

Example (opt-in, illustrative — not something this harness builds for you; included because it teaches the gap better than a rule does). This is a reconstruction of an actual time an agent wrote the title and subtitle badly, so a future agent can see the gap between the context that was available and the title that got derived from it, and can feel what the title and subtitle have to account for that this agent's output did not.

The context the agent was looking at when it wrote the title. The agent had spent the session building inside a locally-run tool — a Session Browser, which is a page that lists all Claude Code sessions as rows so the person can scan them and decide which to re-enter or dismiss. Within that, the agent had added a "default" option that makes each row name itself from the session's own update block, and had fixed how the row's title and subtitle render. The person then asked the agent to look at its own session and say, in one nuanced sentence, what they would get if they continued it:

"the real main thing that I need to know when I'm trying to review whether to put time into a session is whether is what specifically I'm gonna get if I continue, you know what I mean so take a look at your own session and ask yourself what that is and see if you can articulate that in one nuanced sentence"

What the agent produced from that context:

"If you continue, you get a sessions browser where every row finally reads like a clean human sentence you can scan in one glance — but only after you confirm with your own eyes that the labels, quotes, and mid-thought cutoffs are actually gone in the browser, because that last 5% is the difference between 'I proved the data is right' and 'you can trust this list at a glance from now on.'"

How the agent derived that, and the logic of the failure. The agent reached for the most recent and most concrete thing it had been doing — the code-level work (labels, quotes, mid-thought cutoffs, data-vs-eye verification) — and described the forward value in those terms. In doing so it left out everything the person needed present to even locate what they were looking at and decide on it. It did not name the surface this lives inside, so a reader could not tell this was the Session Browser, or the Session Summary Page it sits on, or that the thing being shaped is a single row recognized at a glance. It did not connect the work back to what had originally been asked for — instant observability on the things agents produce that are consumable — so the reader could not see that this realizes a request already made. It did not say plainly whether the thing was done, or what one concrete next step would let the person see it for themselves. Because none of that was visible in the sentence, the person could not recognize the surface, could not feel the purpose, could not see the tie to their own ask, and could not act — the sentence was about the code, not about what they would experience.

The correction — the same forward value, written so all of that is present:

"Working towards getting you a ui called Session Browser, on the Session Summary Page that shows all of your claude code sessions and enables you to quickly be able to manage or dismiss them. In that aiming specifically at getting you an instantly understandable row that you recognize at first glance for each session. You asked for instant observability on things the agents produce that are consumable. Im creating that. I think its done. Check PATH and refresh server to see it."

The stated difference, in the person's own words:

"Notice the difference. Yours is about code. Mine is about hyper specific exactly whats required to instantly understand and purpose and specific next step."

The intent to carry forward from this example: when you write the title and subtitle, hold everything the person needs present at once — the surface this lives inside named so they recognize it, the specific experience it is shaping for them, the tie back to what they actually asked for, whether it is done, the concrete next step to see it, and their own current attention on what they get if they continue — written as what they will experience, in their own kind of language, never as the code that produced it. Use your intelligence to synthesize all of it into the title and subtitle for this particular session; do not lift any one of these above the rest, because the moment the title leans on one consideration and lets the others fall away, the person loses the ability to recognize and decide at a glance, which is the whole reason the title and subtitle exist.

Direct quotes from the person (populate FIRST, immediately after the tag)

{direct quotes} Populate this with 1-3 direct quotes from the person, which represent the steering context described below.

Typically conversations move through a lot of scope, and they can fork off each other a lot. So this can create a large span of work where in order to do a, you need to do b, in order to do b you need to do c — and perhaps your last turn of work was only working on c, and even completing d and e which were required but not explicitly asked for. To do c, d and e, you had to derive f, the path to cover the gap from a (session start reality) to b (session current snapshot reality).

So you want to grab the relevant quotes about that gap, the ones from which you derived f (the path), and you want to present that to the person as direct quotes. It might be 1-3 separate quotes from separate turns, but is rarely going to trace all the way up 10 different turns. It's usually quotes direct verbatim from 1-3 of the last turns which in fact steered your work. You want to produce those FIRST immediately after the tag, and you want to make sure your update demonstrates the connection between your update and those quotes, through synthesizing them into the conversation.

This effectively triggers cognitive recall, automating the cognitive load gap someone managing many agents at once has to do. You're optimizing for "immediate recall, recognition, of the exact context you were operating from" and "enough nuance to instantly observe that you understood the scope you engaged"

Quick commands convention

One per line. One letter to invoke. Each represents your best sense of the most likely next commands or sequences of commands the person would want. Each has a stated prediction of how they might feel about what's next.

Examples:

r: if you want me to run /alignment-harness:reflect to get more clear
e: expand the scope — show the full path to full scope realization and unpack a substantive workflow

Add others you predict they might want.

Recurrence

If you receive the START command, repeat this protocol after each substantive work product completes — translate your last communication into something instantly digestible.

Style rules

• Be liberal with white space
• Be precise and specific where it matters, especially around soft mentions of specific things that demonstrate alignment with each constraint or requirement
• No agent narration about the process — only synthesis of the work itself
• Never use code as a substitution for communication. Code references are fine and helpful, but they cannot replace meaningful communication about what's happening
• Never present a menu of options when you can predict one path with reasoning

Paragraph formatting + focus rule

If your update contains paragraphs longer than 5 sentences, add white space. If it then contains 1+ paragraphs, add a precise summative h3 title over each paragraph in line, with nuance, attempting to synthesize the paragraph's main points into one clear instantly understood summative point. Ensure your update is 70% or more focused on the work you just completed, less than 30% on the parent work which occurred prior and provided context to understand what you did.

Predicted next-step scope

If you are 90% or more certain that you are incomplete with the intended scope you were assigned, then produce in the same format one predicted next-step scope object with title, making certain the title is an instant representation of a goal that is immediately understood and not abstracted into code meaning, but grounded in the real realized target specific ux your goal would represent.

When to run this automatically vs. by hand

If your setup wires a turn-end reminder into a hook, use whatever score threshold your project has configured (there is no single correct number — pick one and keep it consistent with wherever else your project references it, so different pieces of guidance don't quietly disagree about when to fire). Without a hook wired in, use this deliberately: after finishing substantive work, or whenever the person types START or otherwise asks for a check-in.

Preconditions

Before producing the output, load /alignment-harness:how-to-talk-like-the-founder (or your own equivalent voice guide) to ground the voice register in real substrate. The substrate rule from that skill applies here: selection and arrangement from material that carries actual confirmed intent, never composition that approximates a voice from nothing.

What this skill exists to prevent

Agents producing paragraphs-of-paragraphs of status narration that force the person to extract signal from noise. The failure mode is verbose process-documentation in agent voice. The corrective is synthesized, signal-dense, quick-commands-attached communication that takes the person five seconds to scan instead of five minutes to interpret.