Friction report, week 39

This page counts friction in your own Claude Code transcripts and turns the repeats into proposed CLAUDE.md rules. It is for you, so you can rule on each rule. The first count closed on the evening of September 26. An addendum covers the twelve hours after that, so the numbers here run slightly ahead of the markdown you read on Friday.

67friction moments

243messages you typed

28%share of your messages

5rules proposed

How the count was made

Source: every Claude Code transcript under ~/.claude/projects/ touched in the last 15 days, across all project folders and Orca worktrees.
Included: only messages you typed in interactive sessions. Scheduled and headless runs were excluded, which removed all 295 myKG channel-watch sessions.
Excluded: messages from one Claude session to another, pasted text, and resumed-session summaries.
Judging: two agents read every message, one per week, on the same rubric. A word such as "don't" or "always" inside a new instruction did not count as friction. Every quote was then checked against the transcript by hand.

The week in one chart

Counts are friction moments, revised to include the addendum window. Gold marks a pattern that did not appear the week before.

Status chasing on long builds16

Heavier path than asked9

Session cut off or connection lost7

Claimed done, not visible to you6

Billing refusals on outside APIs5

Prose caught after the gates passed3

Retry with nothing changed3

Absence claim never checked2

also present last week new this week

The mix changed more than the rate. Last week, 14 of 18 friction moments were about prose. This week, 5 of 67 were. The writing gate went live on September 16 and 17, which fits the drop, though one week does not prove it caused the drop. What is left is operational: chasing status, checking claims, and scope.

The patterns

1. You had to ask for status on long builds

16 moments · each minor, together the largest cost of the week · report builds

In one session you asked where the aTwist report was ten times between 11:37 and 15:37, roughly every thirty minutes.

"Where is the report at, it's been running for many hours"

"Where is report at?"

"Where is report"

Root cause: Claude reports progress only when you send a message or a background task finishes. A multi-hour workflow has many phase changes that nobody is told about. The live viewer and the rulings log both existed, but you had to know to open them.

2. Claude said something was done that you could not see

6 moments · moderate · settings, Google Sheets, a deployed page, a phone

"memx is blocking again"

"i dont see that tab" … "not showing, take over browser and see for yourself"

"how come the standard view is x'd out and doesnt work? only ambient view shows"

"It doesn't run when I hit run on mobile phone through tailscale link"

Root cause: each claim was checked through an API or a file, not in the view you use. Standing rule 2 in your global CLAUDE.md already says to verify on the real surface. It did not hold for sheets, for settings that need a restart, for interactive toggles, or for a phone.

3. Claude took a heavier path than you asked for

9 moments · minor to moderate · mixed

"actually, just fix memex"

"i mean tell me"

"I don't want to have to tell you to apply the rulings, instruct the agent for me"

"before we start, finish what we were working on, setup the scheduled compute run"

Root cause: the vault's process machinery is visible to every session, and Claude reaches for it by default. A direct request gets a full process.

4. A report said a platform presence did not exist when it did

2 moments, same report · major · the claim was wrong and a search disproves it

"I know there's a tiktok the report says there isnt i mean, a simple google search, for atwist vertical drama tik tok it's the first listing"

"also fix the report, as there are obviously x posts about atwist"

Root cause: the fact-check gate verifies claims that cite a source. A claim that something is absent cites nothing, so nothing checked it.

5. Billing refusals on outside APIs took hours to clear

5 moments · moderate, one major · Jev and OpenAI

"i switched on auto top up and added $5 try again"

"Try again I added $"

"im going to swap out the api key, give me the text to use"

Root cause: mostly outside Claude. TypeSafe refused calls while its console showed credit, and the OpenAI key belonged to a personal organization with no balance. Claude read the organization header only after the second top-up failed. Reading it on the first refusal would have saved a round trip.

The twelve hours the first count missed

The first count closed at 5:27 pm on September 26. You kept working until about 1 am. Thirty messages fell in that gap, across three sessions: the Jev cartography orchestrator, the sources worktree, and the site worktree. Seventeen of the thirty were friction.

That share, 57%, is not comparable to the 28% above. This was a late-night debugging and demo stretch on a half-built system, which is where friction concentrates. What matters is which patterns came back, not the rate.

New in this window

A retry loop. "try again" three times in three minutes, with nothing said about what differed between attempts.
A question you had to ask twice. "What's a good example state I should paste in to try this", then the same question again two minutes later with a second screenshot.
A prose relapse after the gates passed. "It reads like a cheesy pirate children's book. The date was blah blah, and on that date, a system was blah." The rule against opening a sentence with a date went into memory the same minute. It is the first prose correction since the writing gate went live, and it names a failure the gate does not check for.

What worked

The Jev cartography worktree run: five worktrees, a merge, and two audit rounds with no correction on substance.
The Cloudflare Zero Trust walkthrough: eight screenshot-guided steps, no correction.
The Champ School demand scout: "everyone is very happy with this", then only requests to add things.
Deploys: no account, branch, or verification mistake in either week.

Compared with last week

Pattern	Sept 12 to 19	Sept 19 to 27	Trend
Prose you caught after the gates passed	11	3	Down
Numbers or lists written as prose	2	1	Down
Status chasing on long builds	0	16	New
Heavier path than asked	0	9	New
Claimed done, not visible to you	0	6	New
Absence claim not checked	0	2	New
Unglossed internal terms	1	1	Same

The last analysis turned 19 repeated writing instructions into gate rules on September 16. Prose friction fell from 11 to 3 in the week that followed. Two memory rules added this week have not been tested yet: labelled bullets, and no internal names for a client reader.

Last week had fewer build-status moments partly because fewer long workflows ran. The status pattern may be new work rather than a regression.

Proposed rules

Five new rules and one change. Nothing here has been applied. Mark each one approve, change, or reject.

1. Report progress on long builds without being asked

Evidence: 16 status check-ins · Where: project CLAUDE.md, new section "Long-running work"

- When a build or workflow will run longer than 20 minutes, say at launch when it
  should finish and where to watch it. Then use Monitor on its log or status file and
  send Jonny a one-line status at each phase change. He should never have to ask
  where a run is.

2. Take the lightest path that answers the request

Evidence: 9 moments · Where: project CLAUDE.md, "What a good change looks like here"

- "Fix X" means fix it; "tell me" means answer in chat. Start a worktree, workflow,
  or diagnostic plan only when Jonny asks for one or the task cannot be done without
  it, and say which before starting.
- When Jonny rules on something a running agent is waiting for, send the ruling to
  that agent yourself.

3. Check an absence before printing it

Evidence: the report claimed no TikTok and no X presence, and both came up in a plain search · Where: the fact-check stage of /shuriq-report

- A claim that something does not exist (no account, no coverage, no filing) needs a
  logged web search for it, cited like any other claim. The fact-check gate fails an
  absence claim that has no search behind it.

4. Read the account on the first billing refusal

Evidence: five retries over about six hours on Jev and OpenAI · Where: project CLAUDE.md, "Local gotchas"

- On the first 402 or 429 quota error from an outside API, read the error code and the
  organization and project headers, then tell Jonny which account to fund and on which
  page. Do not retry until he says the balance moved.

5. Say what changed before a retry

Evidence: "try again" three times in three minutes · Where: project CLAUDE.md, "Local gotchas"

- When a command fails and Jonny says "try again", name the one thing that changed
  before running it again. If nothing changed, say so, and say the retry only tests
  whether the outside service recovered.

One rule to change

Current rule	Proposed change	Reason
Global `CLAUDE.md`, standing rule 2: "A fix is fixed only when the real surface says so: run the program, curl the deploy."	Add: "For a sheet, a setting, or a control on a page, the real surface is the view Jonny uses. Open it in the browser, click the control, and say what you saw. If a change needs a restart, say it is not live yet. If the view is his phone, the check has to happen on his phone or be declared unverified."	6 moments this week: the Memex setting, the Sheets tab twice, the Standard view toggle, the worktree app inventory, and the Tailscale page on mobile.

Nothing to remove

No existing rule caused friction this week.

Deliberately not proposed

Session cutoffs, 7 moments. Three sessions dropped within seven minutes on September 23, which looks like an outage. A rule would not prevent it.
Bullets, and internal terms in front of a client reader. Memory rules for both were saved this week. Next week's report should show whether they held before anything moves into CLAUDE.md.

Focus for next week

Status on long builds. 16 moments, the largest total cost.
Heavier path than asked. 9 moments, spread across the most sessions.
Checking claims in your view. 6 moments, and it breaks a standing rule you already wrote.

Sources: system/friction-reports/2026-W39-friction-report.md and system/friction-reports/2026-09-26-proposed-rules.md. Every quote is one of your own messages, pulled from your own transcripts. Counts on this page match the markdown. Next analysis due Friday, October 2.