most product teams think they're aligned. they're sharing the shape of an idea, not the shading: the reasoning, the tradeoffs, the context behind every decision. seam transmits the shading.
jeff patton noticed something teams keep relearning: you can put the same shape in front of a room full of people, ask if they agree, and everyone says yes. they do agree, on the shape. what each person imagines inside that shape is different. the artifact aligns. the reasoning behind it doesn't.
AI has made this faster and less visible. your team gets to the same shape in hours instead of weeks. brief, PRD, tickets, all generated, reviewed, approved. the tools agree on every word. what they don't carry is the shading: why this decision over that one, which customer made it matter, which tradeoffs were real and which were just noise. that stays in someone's head, or buried in a thread from three sprints ago.
seam maintains the shading. a shared background that every AI tool in your system reads before it starts. the reasoning behind every decision doesn't get left behind when the feature ships.
when AI is doing the building, context drift doesn't compound over quarters. it compounds over days.
starter kit. free on github. 15-minute setup. it's a set of claude code skills and a short config file. no new server, no new app, nothing to maintain beyond a few text files. ships with the core context layer and the update loop. any team running claude code can install it today.
full configuration. a 4-6 week engagement i run with teams. connects to your tools, sets up the skills library, tunes the shared background to how you actually work.
the feature list below spans both. the starter kit covers the core; the engagement layers on everything else.
open the last three PRDs your team shipped. for each one, can you reconstruct why it was prioritized over what it displaced? if not, that's the shading that didn't travel.
before the feature list, an artifact. this is what one looks like when seam is running.
the weekly signal digest. it gets generated, not written. it pulls from the shared background each monday and posts to slack.
example below is from tinypauses, an indie project i run. tiny mindful moments for kids 9-12 and their grown-ups: short prompts, gentle tracking, no streaks, no pressure. numbers are illustrative.
weekly signal digest | tinypauses | mon apr 21
---------------------------------------------------
SHIPPED
three new "notice" prompts (color, sound, texture).
friday. kids picked "notice a color" twice as often
as the other two over the weekend. check-in: may 5.
dashboard label swap: "your progress" to
"your moments." wednesday. one grown-up emailed
to say it felt warmer. no other feedback.
IN FLIGHT
optional weekly reflection for grown-ups. copy draft
with kate. paused pending one more interview round
with parents who opted in last month.
low-energy prompt set ("i'm not up for much today").
in editorial review. internal concern that the
framing might read as permission to skip rather than
a genuine alternative. decide this week.
USER SIGNAL
7 parents replied to the daily email this week.
3 asked for bedtime-friendly prompts. current library
assumes daytime. not yet a ticket.
1 teacher replied asking if there's a classroom
version. 2nd time this has come up in 6 weeks.
custom territory. logged, not acted on.
daily email open rate steady at 34%. click-through
to the dashboard dropped from 18% to 12% over three
weeks. possibly the new email layout. eng looking
at it.
FLAGS
one parent flagged "notice what feels heavy in your
body" as uncomfortable for an anxious 9-year-old.
reviewed with kate. prompt stays, gets an opt-out
alternative. logged apr 18.
low-energy prompt set: two weeks in review is the
ceiling. ship, revise, or pull this week.
DECISIONS THIS WEEK
apr 16 declined a "pause of the day" badge on the
dashboard. reason: even a soft badge
re-creates the streak pressure we
specifically removed. revisit only if kids
ask for it directly, not grown-ups. owner:
maya.
apr 18 kept email over push for the daily prompt.
push has better open rates on paper but it
felt too "performance app" for the 9-12
audience. decision reviewed, email stays.
WHAT'S QUIET
zero replies to the "notice a sound" prompt this
week. not a problem yet, but worth watching. if
silent for another two weeks, revisit the copy.
no one wrote this. the shared background already knew what shipped, what was logged, which signals repeated, and which dependencies were aging. the digest is a read-out, not an artifact anyone had to sit down and produce.
the same logic produces every other artifact on the list below. the PRD knows why it's being written. the triage knows what's already decided. the brief knows who it's for. no one re-explains the company at the top of every document.
pull up the last AI-generated artifact your team produced. could a teammate read it cold and tell which customer it's for, which goal it serves, and which decisions shaped it? if it reads generic, your context layer is thin.
the starter kit covers the core context layer and closes the feedback loop. skills update your context files automatically after each run. the full setup connects to your tools and surfaces signal on its own.
each one reads from the same context layer. write the company once; every skill inherits it.
without a shared context layer, pacing zone classifications drift because the context drifts. a surface that was zone 1 in march is zone 3 by june because a new dependency landed, a compliance exposure showed up, or a customer segment changed. teams end up re-litigating zone calls every sprint, running the classifier against whichever version of the context the loudest voice holds. seam is what keeps the classifier honest. when the shading updates, the zones update with it.
none of this replaces strategy. but when the shared background in seam holds current goals, customer signal, and in-flight work, pacing-zone calls stop being a gut read. the classifier runs against your vision, your strategy, and your actual goals.
the "did it work?" check closes the loop. the outcome review produces a summary of what the team learned, which updates seam. every AI tool reads from the updated shared background on its next run.
how the shared background keeps itself up to date. this part is live today.
the "did it work?" review produces a summary of what the team learned. one person reviews and approves. seam's shared background updates. every AI tool reads from the updated background on its next run.
the weekly digest feeds into this too. customer signal that gets confirmed or disproved flows back into the shared background automatically. no one maintains files by hand. the loop writes them.
seam handles alignment inside a team. the harder version, between teams and between companies, is still open.
when you send a brief to a client or partner, your AI transmits the shape: the document, the proposal, the summary. the shading doesn't travel: the reasoning, the tradeoffs, the customer context that made every word choice deliberate. their AI reconstructs it from scratch with different goals and different priors.
today, the workaround is manual. context export packages the relevant background from seam into a structured format the recipient can actually use. one human reviews before it leaves the team.
what's missing for cross-team context at scale: the decision layer. who decides what gets transmitted, under what conditions, with how much human judgment in the loop. the technical groundwork is there. the judgment layer isn't.
seam is that layer in its first form. today, inside your team. what comes after that is still being built.
the other open problem: context decays. shading that was accurate six months ago may not reflect where the market or the team is now. the weekly signal digest surfaces when context has drifted. gate 3 review caps close the loop (up to 14 days for zone 1, 30 for zone 2, 60 for zones 3 and 4, with the review firing when signal lands). a more formal re-validation mechanic is on the roadmap.
built for product and support teams at B2B SaaS companies. the ones where being close to the customer is what differentiates them, but the infrastructure to act on that proximity doesn't exist yet.
when seam is running, the team isn't the bottleneck anymore. support has a clear channel and knows what to say. product has a live picture of what customers are asking for and how it maps to the roadmap. engineering ships toward goals instead of toward whoever asked last. and the AI tools everyone uses are working from the same shading, because they're drawing from the same source.