From an engineering manager skimming titles and captions:
- "but a check after that fails" under "Mission complete!" where the
prompt said that step may fail now adds ", as you said it might".
- Saves no prompt of the turn asked for (a project rule, or Claude's own
idea) were indistinguishable from requested ones: the turn's first such
save carries a small line, "a save your prompt here did not ask for".
"save its log with tee" no longer counts as asking to save the work.
- The first save on a branch Claude created says which: "saves the work
on a new branch (feat/csvstat)".
- "Claude changes 3 files" hid an edit to your CLAUDE.md: a panel that
wrote some files and patched others says both ("Claude writes
README.md and .gitignore and changes CLAUDE.md."); a file created and
then edited in the same panel is written.
- A turn with a problem left and work still out read "Done for now";
the catch now wins the title.
- A helper's report that is only a path reads "reports back with the
file it wrote: README.md."
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From the fact-checker:
- "+37" sat on mathx.py's card, which has 11 of those lines; with
several files the badge now counts the card in front (the Director's
cut keeps every file's count).
- A second opinion asked in the middle of a step's calls read "after
asking"; it now reads "along the way, asks a stronger AI model…".
- "Two places where I didn't follow your wording exactly" under
"Mission complete!": not following your words is a note ("Done, with a
note"), and the sentence counts as a word on what you asked.
- A test run that patched a file written a panel before said "Claude
writes" it: whole files are written, patched ones changed ("Claude
writes pyproject.toml and changes test_csvstat.py; …").
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From the person who asked and a fact-checker, reading fresh pages:
- "as planned" for failing first runs the prompt asked for ("it should
fail because mode is missing", "run them to watch them fail"): those
read "The first run fails on purpose, as you asked".
- A repository Claude started was told only on a save panel; the first
save inside a loop or a test run says it too ("Claude starts a git
repository and saves the work in 4 commits along the way.").
- `git push && git describe --tags; echo "exit=$?"` exits 0, so a
failure after the push went unsaid; what it printed after the push
(`fatal:`, `exit=128`) now counts. The commit subject stays on that
caption too.
- Tool-call counts no longer include second opinions.
- "a layout check" is "a formatting check"; "Turn over" (a game word)
is "Not done".
- "the last check finds nothing" told nothing: it now says what came
back empty ("the last one finds a git setting that is not set, which is
no error").
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From the person who asked for the work, reading fresh pages:
- "The result was 4 passed, 1 failed" under "Mission complete!" read as
a failing end. A test count is no failed step, and Claude's second
sentence is kept when it is the outcome ("All 5 tests now pass, and the
change is committed.").
- "Two places where I didn't follow your wording:" showed one. A list
intro that announces how many brings that many lines.
- "…would fix it for up to 10 files. Want me to add that?" never said
what "it" was. An offer that points back brings the start of its
paragraph ("Suffix swap: the two files with the same time would trade
their _1 suffix. …").
- "Decisions you may want to change" is something for you, like an offer.
- A second opinion whose call failed read as given; it now reads "after
trying to ask … (no answer came)".
- "but a step after that fails" after a push sounded as if the push
broke: when everything after it only looks (git log, status), it is "a
check after that fails".
- Tests run in a fresh clone read "The tests pass in a fresh copy made
from remote.git."
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 9: a background job's page showed only Claude Code's
line 'Agent "Helper B: write extra pytest tests" finished', so what the
helper said about its own work vanished, here that it had never run its
tests against the real wordcount.py.
The notice balloon now carries the helper's report like Claude's last
words (its first sentence and the key sentences from the rest):
'"Helper B: write extra pytest tests" reports: I wrote the file:
tests/test_extra.py It has 18 test cases … I haven't run it against the
real wordcount.py.' A report that is only a path is its file name, and
absolute paths keep their last two parts, so a long path does not push
the sentence that matters out of a phone's balloon.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 9: "You stop Claude here, during its step 3." said
where Claude was, not what that meant. The stop now carries the step's
own words, "during its step 3 (new tests under test/)", and when Claude
kept a to-do list, what was still open on it: "Not done yet: Write the
new tests; Update the README and 1 more."
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 9 (a newcomer to the project):
- One or two files are named, not counted: "Claude writes .gitignore and
saves the work", "Claude changes rename_by_date.sh.", "Claude writes
the new files mathx.py and test_mathx.py." Three or more are counted.
- The first save in a folder Claude turned into a repository says so:
"Claude starts a git repository and saves the work: …". A `git init
--bare` (the push target) does not count, nor a folder that already
had commits.
- A push to a repository in a folder on this computer (`To ./origin.git`)
is a copy, not an upload: "then copies it to origin.git, a folder on
this computer", and the final chip reads "copied to origin.git".
- A tool lookup for plan mode reads "Claude gets ready to plan."
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 9:
- "another check" was too vague to act on. A check other than the tests
is named by its tool: a layout check (a formatter), a style check (a
linter), a type check, the build. "Fix and test, 2 rounds: a test fails
and Claude fixes it, then a layout check fails and Claude fixes that;
everything passes."
- A failure the prompt asked for reads "The first run fails on purpose,
as you asked; the second passes." ("The tests don't even load, as you
asked."); one Claude planned itself keeps "Test first: … as planned".
- A test run sent to the background that reported in a later turn reads
"Claude starts the tests in the background; later they pass.", not as
if they passed in that turn.
- The test board says "1 test" and, when nothing loaded, "1 error".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Two fresh cold readers (the person who asked for the work, a newcomer to
the project) found the round-8 balloons still cut exactly where Claude
turns from describing to asking, offering or warning: "Want me to add
that?", "I can fix the number format whenever you want", the
http.server instruction, the fix that followed a reported bug.
Key sentences are now picked by kind, one of each, at most three, and
shown in the order Claude wrote them: a problem left (with the fix right
after it, when there is one, so it never reads as open), a question or
offer, how to use what was built, something not done or checked yet, a
step that failed, a word on what you asked for, a corrected warning. A
list intro is judged on its own words and brings its first items; a
pointer ("described below") is skipped; a short offer brings what
"that" was.
The final panel reads "Done, with a catch" for a problem left, "Done,
with a note" for something pending, and "Done, with a question" when
the next move is yours (`openNote` is now 'catch' | 'note'; `asks`).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 8: you asked to see a dry run, Claude's balloon said
"Here's the dry run you asked for", and the page showed nothing of it.
A run panel's terminal card now prints, dimmer under the commands, the
first lines its last command printed (the last ones when it failed:
that is where an error is).
The card is drawn on the canvas, so it goes into a saved image someone
may share: a line that looks like a secret (token, password, api key,
credential, inside identifiers too) is left out, addresses and long
token-like strings are masked, and Claude Code's own markers ("(Bash
completed with no output)", tool-error tags) are dropped.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 8: "STOPPED" with the balloon "Step 3: new tests under
test/." read like progress, and nothing said what the stop left. The
final panel now names the step Claude last announced and the commits made
before the stop: "You stop Claude here, during its step 3. Saved before
the stop: 2 commits."
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 8: "Mission complete!" with confetti and cheering
helpers sat over balloons saying the page will not work as opened, a
helper's tests have not run, or a problem is left. The final panel is now
calm (no confetti, a neutral robot) whenever work was still out at the
turn's end, nothing worked, or Claude's last words name an open problem
(`page.openNote`: a key point about something wrong, not done, not
checked or not fixed); the last of these reads "Done, with a catch".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Two agents read only the comic of the eleven recorded sessions (one who
has never written code, one engineering manager skimming), wrote down
what they believed, then compared with the transcripts. Both were misled
in the same places, and the cause was the same: the final panel kept
Claude's first sentence or two, and what a reader must not lose sat
further into its last message.
- Claude's last words now keep key sentences from anywhere in its final
message (read in full, past the label's clip): an open problem or a
quirk, something not yet done, checked or confirmed, work still out,
what you asked for or expected, a correction of an earlier warning
(with the sentence it refers to). A bare "I didn't" counts only when
nothing stronger is there; table rows, list intros and questions never.
"I built the pomodoro timer. Opening the page: it won't work if you
double-click index.html…"; "…That note makes the current behaviour
clear, but it doesn't fix it."; "As you expected, git describe --tags
then failed because the repo has no tags."
- A turn that ended with helpers or a background command still out says
so: "Done for now", and "Claude's part is done in 36 seconds. Its 2
helpers finished later and reported back." (`finishedLater`).
- A failure you asked for, or that Claude said it expected, reads as the
plan: "Test first: the first run fails as planned, the second
passes.", "The tests don't even load, as planned." A loop that failed
once on a test and once on another check says both, and the commits
made inside a loop are part of its story; the final panel counts the
turn's commits ("5 commits, uploaded").
- One or two files read are named ("Claude checks README.md." in a
turn a report started); a tool lookup reads "Claude gets the tools it
needs ready." again (it was overwritten) without its query; the
advisor is "a stronger AI model", and one that answered with an error
"gets no answer"; a commit message loses its `feat:` prefix in the
plain caption.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Round 6's two sessions join the corpus (sanitized like the others, and
checked to draw the same pages as the originals): tests that do not load,
a fix-fail-fix-pass loop, a commit then a push in its own command, a push
with a failing step after it, an append; two helpers and a command in the
background whose reports come back as turns of their own.
Replaying them showed two captions to fix:
- A test command that ran in the background and reported back read "The
tests pass.", as if they ran during that turn: "The tests run in the
background and pass." (events keep `background`).
- A python script that rewrote README.md drew as "Claude works in the
terminal." beside a separate change panel. A plain command that writes
project files (a script fed by a here-document, or one Claude Code says
changed files) is an edit: "Claude changes 3 files."
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review round 6: commit, then push as a second command (the most
common way to ship) drew as "Claude saves the work in 2 commits and
uploads it." for one commit, and the commit's message was dropped. A
command's `commits` was only recorded above one, and the save panel then
counted every save command as one. It is now recorded whenever a command
commits and is the only thing counted: "Claude saves the work and uploads
it: "feat: add mode()"." A push of earlier commits reads "Claude uploads
the work."
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- The bar on a phone took four rows (216 px of a 664 px screen) with the
close button alone on the last; it now takes three (170 px): title and
close, the turn, the view and Save image. The drawer starts exactly
under it and follows it when the phone turns (the bar is observed).
- A panel below the fold had no buttons until it was scrolled to (only
drawn panels get click regions), so Tab skipped it. A quiet frame now
draws one such panel, a band at a time; its entrance still plays when
it comes into view.
- "Fix and test: the first run fails, the second passes." was chosen by
rounds, so a loop with three runs (a failing test, then ruff, then a
pass) said it under ✗ ✗ ✓. It now needs exactly two runs; that loop
reads "Fix and test, 2 rounds: 2 runs fail, then it passes.", and the
corpus pins every loop's marks.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Review round 5 had an agent read only the comic of the nine real sessions
(20 turns), write down what it believed happened, and only then compare
with the transcripts. Three panels told it something false, one hid the
failure the user asked to see:
- A push that went through, followed by a check that failed in the same
command, drew as a failed push; the output (`[new branch]`, `->`) now
decides: "Claude saves the work in 2 commits and uploads it, but a step
after that fails."
- A formatter's "would be reformatted" preview, printed before it fixed
the file, made a clean run a failure: the last verdict wins.
- A blocked command ended in confetti and "Mission complete!": a run panel
where everything failed says "Claude tries a command, and it is
blocked.", and a turn where nothing worked ends "Turn over" / "Nothing
Claude tried worked, so the job is not done." with a neutral robot.
- The loop stamped only the final verdict (POW ×2), so a first run that
failed on purpose never showed. It now marks every test run red or green,
and says it: "Fix and test: the first run fails, the second passes." /
"Build and test, 4 rounds. A check fails once, then everything passes."
Also:
- Files written inside a check or a save reach the plain caption ("Claude
writes 4 files; the tests pass.").
- Balloons keep the next sentence when the first reports a problem ("…left a
stray file. I'll remove it…") or is very short ("Not done. The harness
blocked sleep 65…"); Claude's last words keep a caveat that follows them
("Nothing is committed because this folder isn't a git repository.").
- The advisor never shares a panel with helpers that do work; ToolSearch
reads "Claude gets the tools it needs ready."; Codeman's own case file is
not drawn as a file Claude read; looking around shows what Claude meant
(the call's description), not raw commands; a montage at the start of a
turn reads "Claude gets started".
- Words: boards count tests ("19 tests"), "uploads it" replaces "sends it
up", and the prompt balloon keeps 400 characters.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Nine transcripts from the review rounds, each a real Claude Code worker
building a small program (a Python CLI with a fix loop, a web page with
node:test, a bash script with a background helper, a long turn with a
prompt added mid-turn, teammates then an interrupt, a question then tests
first then a push, a text plan with a scratch dry run, /compact then plan
mode, a test failed on purpose). Each is cut down to the lines and fields
the reducer reads, with every path, user name and email rewritten and the
file contents of Reads left out; replaying them before and after gives
the same pages.
The expected file beside each pins every page (panel kinds, captions,
balloons, counts), so a change that alters a story fails with a diff of
exactly what changed; STORY_CORPUS_UPDATE=1 regenerates them. The beats
the reviews were about are also stated outright.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>