Audit Claude Code Insights

August 21, 2026 - Audit a Claude Code /insights report and record a verdict.

Introduction

30 Second Summary

A polished report can feel convincing the moment you read it. That first impression makes it hard to tell whether the report revealed something new or simply echoed what you already knew.

In this project, you will audit Claude Code's /insights report to decide whether it deserves a place in your work-review routine. Your final verdict will trace every count to your frozen list or the report itself.

What You'll Build

You'll read your finished worksheet and see a dated trail from your untouched prediction to an Adopt, Reject, Watch, or Park verdict backed by evidence from your own report.

By the end of this project, you'll have:

A sealed prediction you can use to prove which work frictions you remembered before the report could influence you.
A four-measurement evidence review that traces session coverage plus every friction count to the frozen list or report.
A time-counted verdict with a machine-only scope, a reversal trigger, and a re-check date when required.
Secret Mission: Audit the most ambiguous match without changing the sealed prediction, then preserve any revised count or verdict as a visible correction.

Are there any prerequisites?

This project expects a Windows machine with Claude Code already installed and signed in through an existing subscription. You also need a default browser plus meaningful local Claude Code session history on that machine.

Before We Start

This first checkpoint freezes when your evaluation began before any report can shape your judgment. It also commits you to deciding whether Claude Code /insights earns a place in your work-review routine.

Verify the Zero-Setup Path

Your evaluation clock is already running. Setup time belongs inside the thirty-minute result.

The existing Claude Code client should provide the whole path on this Windows machine. Verifying it now protects the evaluation from hidden installation work.

Your existing subscription also defines the cost boundary. No new account, metered API key, payment card, or added service enters this project.

In this step, get ready to:

Record the platform and access facts.
Confirm the existing client opens normally.
Preserve the existing cost path.

Record the platform and setup facts

A zero-setup claim needs an audit trail. These facts separate resources you already had from anything introduced by the evaluation.

Return to the project worksheet from Step 0.
Record the platform and setup facts by copying this block:

Run platform: Windows
Installed for this project: nothing
Existing access: Claude Code subscription
New account created: no
Metered API key added: no
Payment card added: no
Browser: existing default browser

Why Record These Facts?

These fields prove that the evaluation started with an existing client plus existing access. They keep hidden setup work out of the final verdict.

Add a note beneath the setup facts that no additional service entered the project path.
Save the project worksheet.

Good, your worksheet now names the exact environment behind this evaluation. The zero-setup claim has a written baseline.

Confirm the existing client

This readiness check tests the client already on your machine. An unavailable command means the prerequisite is unmet.

Switch back to the existing Windows terminal.

Before you run the readiness check, do you expect both commands to use the existing client without changing your setup?

Check both readiness points by running these commands:

claude --version
claude

What Do These Commands Prove?

The first command asks the existing client to print its version.
The second command starts Claude Code with the existing subscription login.

You'll see a version from the first command. The second command opens Claude Code normally in the same terminal.

Is Claude Code Unavailable?

An unavailable client breaks the zero-setup prerequisite. Keep installation work outside this evaluation.

Stop the project clock.
Mark the prerequisite as unmet in your worksheet.
Leave the client setup unchanged for this project.

Help me assess the unmet prerequisite without installing anything.

Record that claude --version succeeded in your worksheet.
Record that claude opened normally in the existing Windows terminal.

Confirm the cost path

The $0 incremental path still consumes your existing plan allowance. The /insights analysis uses the same account as regular sessions.

At this point, /insights stays unrun. The next step must seal the baseline before any report appears.

In the project worksheet from Step 0, record that this run uses your existing subscription login. Record that /insights consumes existing plan usage.
Save the project worksheet.

Before you review the finished setup record, do you expect to find any new installation or access path?

Review the setup section of your worksheet against the terminal check.

You'll see Windows as the run platform. The setup facts show nothing installed, existing subscription access, no new account, no metered API key, no payment card, no added service, and the existing default browser.

The readiness note says claude --version succeeded. It also says claude opened normally in the existing Windows terminal.

The cost note says /insights consumes existing plan usage. The $0 incremental claim stays limited to added cost.

That is your clean baseline: the existing client works without changing the project's access or cost path. Next, you'll seal the pre-run record that protects the comparison from hindsight.

Freeze Your Decision Record

Your zero-setup check proved that the existing Claude Code client is ready on Windows. The evaluation still needs protection from hindsight bias before any report appears.

A sealed prediction fixes what you remember before /insights can influence it. You will also lock the method that turns those memories into a verdict.

In this step, get ready to:

Complete the pre-run decision record and execution plan.
Preserve the exact dated publisher anchor.
Freeze a numbered memory-only friction list before running /insights.

Complete the decision record

The decision record establishes your evidence order before the report can shape your expectations. It keeps the eventual verdict tied to a method you chose in advance.

Return to the project worksheet from earlier.
Add the pre-run record by copying this template:

Decision question: Does Claude Code /insights earn a place in my work-review routine?
Planned method: Freeze recall, run /insights once, capture the header before findings, transcribe report-named frictions, compare one to one, record four measurements, then write a verdict.
Evidence rule: Numbers come only from the frozen list or report.
Decision rule: Counts inform the verdict, but the comparison is a judgment rather than an equality.
Routine under consideration:
Reversal trigger required: yes
Re-check date required for Watch or Park: yes

How Does This Record Protect the Decision?

The planned method fixes the order in which you collect evidence.
The evidence rule gives every later number a traceable source.
The decision rule leaves room for practitioner judgment.
The reversal trigger makes the verdict testable in the future.

Enter your named routine after Routine under consideration:.
Save the project worksheet.
Scan the decision record for a predicted verdict.

You should find a completed method with no predicted verdict. Your final decision remains open until the evidence exists.

Unsure How to Name the Routine?

Name the real review habit you are evaluating. A weekly work review or quarterly workflow audit gives the later verdict a concrete destination.

Need help defining a routine that fits your work?

Add the dated publisher anchor

A dated publisher anchor separates the verified product record from your later judgment. The anchor establishes which release wording supported this evaluation.

Add the publisher record beneath the decision record by copying these three lines exactly:

Claude Code v2.1.239
Published: 2026-08-21T19:54:23Z
“Fixed `/insights` echoing literal `<analysis>` tags in its response on some models”

What Does the Anchor Establish?

The anchor records the published v2.1.239 fix with its timestamp. Your later measurements determine whether the command deserves a place in your routine.

Save the project worksheet.
Compare all three publisher lines with the reference above.

You should see the version, publication timestamp, and quoted fix preserved character for character.

Publisher Wording Does Not Match?

Copy the three lines directly from the reference block. Preserve the capitalization, punctuation, backticks, angle brackets, and curly quotation marks.

Help me compare my publisher anchor with the required wording.

Write and seal the prediction

Your memory-only friction list becomes the baseline for the later comparison. Keeping examples and previous records out of view prevents them from priming what you write.

Keep transcripts, prior reports, and friction examples closed.
Write a numbered list of specific frictions that recur in your Claude Code work.
Count the numbered items on your list.

That is the baseline captured from your own memory. The next block seals it against later edits.

Place your cursor beneath the final numbered item.
Add the seal by copying this block:

FROZEN
Frozen at: YYYY-MM-DD HH:MM:SS local time
Items on sealed list:
Initials:

What Does Frozen Mean?

The FROZEN marker creates the evidence boundary. Later report findings can be compared with this list without changing what you originally remembered.

Any later correction belongs in the execution record with a new timestamp. The original list stays visible as the baseline.

Record the current local date and time to the second as YYYY-MM-DD HH:MM:SS local time.
Enter the list total as count.
Compare the recorded count with the numbered list.

The two counts should match exactly. That agreement makes every later classification traceable to a sealed item.

Add learner initials after Initials:.
Save the project worksheet.

Before you inspect the finished record, do you expect every timestamp and count to prove that the prediction came before the report?

Check that Frozen at: is later than the existing Project opened timestamp.
Check that Items on sealed list: matches the numbered list.
Check that the exact publisher anchor appears above the numbered list.
Confirm that /insights has not been typed.
Confirm that no insights report has been generated or read.

You should see a complete decision record, exact publisher anchor, numbered friction list, and timestamped seal. The frozen list is now ready for an evidence-first comparison.

Does the Seal Fail a Check?

Correct an empty seal field before continuing. If the numbered list itself has already been frozen, preserve its wording while fixing only the count or timestamp field.

Help me verify my frozen decision record without changing the sealed list.

Your worksheet should now match one of these states.

✔️ My record is frozen

The method, publisher anchor, numbered list, timestamp, item count, and initials are present. Keep the frozen list unchanged when you continue.

ⓧ I'd like to double check the full record

Compare the fixed structure below with your project worksheet.
Keep your own numbered memory-only list between the publisher anchor and the FROZEN block.
Keep your completed values after the routine, timestamp, count, and initials labels.

Run platform: Windows
Installed for this project: nothing
Existing access: Claude Code subscription
New account created: no
Metered API key added: no
Payment card added: no
Browser: existing default browser

Decision question: Does Claude Code /insights earn a place in my work-review routine?
Planned method: Freeze recall, run /insights once, capture the header before findings, transcribe report-named frictions, compare one to one, record four measurements, then write a verdict.
Evidence rule: Numbers come only from the frozen list or report.
Decision rule: Counts inform the verdict, but the comparison is a judgment rather than an equality.
Routine under consideration:
Reversal trigger required: yes
Re-check date required for Watch or Park: yes

Claude Code v2.1.239
Published: 2026-08-21T19:54:23Z
“Fixed `/insights` echoing literal `<analysis>` tags in its response on some models”

FROZEN
Frozen at: YYYY-MM-DD HH:MM:SS local time
Items on sealed list:
Initials:

How Should the Full Record Differ?

Your worksheet includes the existing project-open timestamp from earlier. It also includes your named routine and numbered memory-only list.

The timestamp, count, and initials fields contain your completed values. Every fixed line matches this reference exactly.

Your baseline is sealed before any report evidence appears. Next, you will run /insights and capture its count-bearing header before reading a single finding.

Run Insights and Capture the Header

Your Claude Code decision record is sealed. It now preserves the frictions you remembered before seeing any report evidence.

The report covers a partial local sample by design. Capturing its count-bearing header first preserves the sample coverage before any finding shapes your judgment.

In this step, get ready to:

Generate one local report from the existing signed-in session.
Open the generated report at the top of the page.
Copy the count-bearing header before reading any findings.

Generate the local report

The /insights command analyzes recent eligible sessions on this Windows machine. It writes the result as a local HTML report.

Switch back to the existing signed-in Claude Code session.
Analyze your local session history by entering this command once:

/insights

What does this command do?

The /insights command uses recent eligible sessions from this machine as its source.
Claude Code writes the latest result to report.html.
The analysis uses the provider from your regular sessions.
Its tokens count against your existing plan usage.

Wait for Claude Code to finish writing the report.
Keep the printed report link visible.

You'll see a link to the generated report in the Claude Code session. Good, your evidence source now exists on this machine.

Report link not printed?

Confirm that you returned to the local Claude Code session from earlier. The command is unavailable in cloud sessions.

Keep the session open until the analysis finishes. Very short sessions can be skipped when Claude Code selects the report sample.

Help me diagnose why my local report was not generated.

Open the report without reading findings

The printed link points to the latest local report.html file. The fallback path reaches the same file when the link does not open automatically.

✔️ The report link opened

Follow the report link printed by Claude Code.
Keep the browser page at the top.

You'll see the generated report in your existing default browser. Its count-bearing header should be visible before any findings.

ⓧ The report link did not open

Switch to the existing default browser.
Enter %USERPROFILE%\.claude\usage-data\report.html in the browser address bar.
Press Enter.
Keep the browser page at the top.

You'll see the latest local report generated by Claude Code. Its count-bearing header should appear before the findings.

Report file will not open?

Return to the Claude Code session from earlier. Confirm that the report finished generating before using the fallback path.

Check that the browser address matches %USERPROFILE%\.claude\usage-data\report.html exactly.

Help me open the generated local report.

Capture the header before findings

Evidence ordering keeps the sample definition separate from your reaction to the findings. Preserve exactly what the report says about its session counts.

Keep the report at the top of the page.
Locate the full count-bearing header without scrolling to a finding.
Copy the complete header character for character.
Switch back to the project worksheet from earlier.

Use the tab that matches the count information shown in your report header.

✔️ Both counts are stated

Record the current local time to the second using the same Windows clock method from earlier.
Add the header record to the worksheet by completing this template:

Report header, verbatim: [[REPORT_HEADER="full count-bearing header copied character for character"]]
Header captured: [[HEADER_CAPTURED="YYYY-MM-DD HH:MM:SS local time"]]
Ordering check: Header captured after FROZEN and before findings read

What does this record prove?

The first line preserves the report's analyzed session count plus its total session count. The timestamp fixes when you captured that evidence.

The ordering line records that your sealed prediction came first. It also confirms that no finding influenced the header record.

ⓧ Total is not separately stated

Record the current local time to the second using the same Windows clock method from earlier.
Preserve the missing denominator by completing this template:

Report header, verbatim: [[REPORT_HEADER="full count-bearing header copied character for character"]]
Header captured: [[HEADER_CAPTURED="YYYY-MM-DD HH:MM:SS local time"]]
Ordering check: Header captured after FROZEN and before findings read
Total not separately stated by report

Why preserve the missing total?

The report header remains the authority for every session count. Recording the absent total prevents an estimate from entering the evidence.

This gap becomes part of the final scope limit. Your verdict can reflect uncertainty without inventing a denominator.

Before you compare the timestamps, which record must come first for the sealed prediction to remain independent of the report?

Save the project worksheet.
Compare the FROZEN timestamp with the Header captured timestamp.
Leave the browser report at the top of the page.

The FROZEN timestamp should be earlier than the Header captured timestamp. Your worksheet should contain the verbatim header without any transcribed finding text.

Header or timestamp looks wrong?

Return to the top of the report if the copied header is missing punctuation. Copy it again before scrolling farther.

Use the same local timestamp format from the frozen record. Keep the original FROZEN time unchanged.

Help me verify my report header record without reading the findings.

Your report's sample boundary is now preserved beside the unchanged frozen record. Next, you'll read the findings while keeping this evidence order intact.

Transcribe the Report Findings

You already protected the count-bearing header before reading any findings in your Claude Code /insights report. That ordering keeps the report's sample boundary attached to everything you read next.

A report can feel insightful without proving that it added new evidence. This step lets that first impression stand on its own before you test it against traceable rows.

In this step, get ready to:

Preserve every distinctly named friction in a separate worksheet row.
Test whether your first impression can support a routine decision.

Transcribe each named friction

A distinct friction is an issue the report names separately. Similar themes can still represent separate findings.

Each worksheet row preserves one report claim. This structure makes the later comparison traceable.

Switch back to the existing browser tab containing the report from earlier.
Scroll from the count-bearing header to the report's friction findings.
Read the first distinctly named friction.
Add a separate report-finding row to your existing worksheet.
Copy enough of the report's wording into that row to identify the finding later.
Repeat this row pattern for every distinctly named friction.
Keep separately named findings in separate rows when their themes overlap.
Leave the frozen memory-only list untouched.

What Counts as One Friction?

One report row represents one distinctly named friction.
Overlapping themes remain separate when the report names them separately.
Identifying report wording keeps each row tied to its source.
Suggestions describe possible actions. They provide no proof that a fix happened.

Your worksheet now gives every named friction its own traceable place while preserving the sealed prediction.

Write the unsupported first impression

Your report rows are now complete. The next beat captures your immediate reaction before any classification begins.

Before you write it, predict whether one immediate-reaction sentence can carry a reproducible verdict.

Write one sentence in your worksheet answering whether the report felt useful.
Read that sentence on its own.

The sentence contains a reaction. It has no reproducible count from the report rows or frozen list.

That shortfall is the intended result.

Add this exact label directly below the sentence: UNSUPPORTED IMPRESSION: This is recall, not the verdict.
Keep the first-impression sentence visible above the label.

Why Does the Impression Fall Short?

The sentence records an honest reaction. It carries no reproducible count from the report rows or frozen list.

The label keeps the reaction visible without giving it decision weight.

Before the final check, predict whether the frozen list still matches the sealed version character for character.

Trace every distinctly named friction in the report to one worksheet row.
Confirm that each row retains enough report wording to identify its finding.
Check that no suggestion appears as a completed fix.
Compare the frozen list with the version sealed before the report opened.
Confirm that the label reads exactly: UNSUPPORTED IMPRESSION: This is recall, not the verdict.

You should see one worksheet row for every distinctly named report friction. The exact unsupported label sits below your one-sentence impression while the frozen list remains unchanged.

No suggestion appears as a completed fix.

You preserved each report friction without touching the sealed list. Next, you will compare those rows with the frozen record to turn the unsupported impression into measured evidence.

Count the Evidence and Decide

You have preserved the Claude Code report header before reading its findings. The frozen list also remains untouched.

A useful first impression cannot support a recurring routine decision. This step replaces recall with traceable counts before you choose a verdict.

In this step, get ready to:

Classify each report friction against the frozen list.
Trace four measurements to report rows or sealed-list items.
Complete the final review with timing, limits, and decision conditions.

Classify each row one to one

A Confirmed report row matches one item on the sealed list. A New report row has no sealed-list match.

Each report row can match at most one frozen item. Each frozen item can confirm at most one report row.

Switch back to the project worksheet from earlier.
Compare each transcribed report row with the untouched frozen list.
Label a report row Confirmed when it matches one sealed-list item.
Write the matched frozen item beside that Confirmed row.

Each completed Confirmed row now shows the evidence link behind its classification.

Label every unmatched report row New.
Review each frozen item after all report rows have a classification.
Label every unmatched frozen item sealed-list item not raised.
Add one sentence explaining the judgment beside each classified report row.

The worksheet now separates evidence the report confirmed from evidence it added. It also preserves remembered friction that the report did not raise.

Count the report rows labelled Confirmed.
Count the report rows labelled New.
Count the frozen items labelled sealed-list item not raised.
Add the four required measurements to the worksheet by pasting this template:

1. Sessions analyzed and total: [[SESSION_COVERAGE="analyzed of total, or report absence wording"]]
2. Frictions confirmed: [[CONFIRMED_COUNT="number of confirmed frictions"]]
3. Frictions new: [[NEW_COUNT="number of new frictions"]]
4. Items on sealed list the report did not raise: [[NOT_RAISED_COUNT="number of sealed-list items not raised"]]

Where do the measurements come from?

Measurement 1 comes directly from the verbatim report header.
Measurement 2 counts report rows classified as Confirmed.
Measurement 3 counts report rows classified as New.
Measurement 4 counts unmatched items on the frozen list.

Complete all four measurement lines using the evidence you just counted.
Enter Total not separately stated by report when the header does not provide a separate total.
Save the project worksheet from earlier.

Before you check the totals, do you expect every report row and frozen item to appear exactly once in the arithmetic?

Add the Confirmed count to the New count.

The result should equal the total number of transcribed report rows.

Add the Confirmed count to the sealed-list item not raised count.

The result should equal the item count recorded beneath FROZEN.

Counts do not reconcile?

A mismatch in the report-row total means a row has no classification or received two classifications. Return to the transcribed findings to locate it.

A mismatch in the frozen-list total means one sealed item was matched twice or left without a status.

Help me audit my one-to-one classification counts.

Reconcile the plan with the run

A reproducible audit preserves any difference between the planned method and the actual run. Corrections sit beside the original record so the evidence history stays visible.

Compare the pre-run plan with the execution record.
Append a timestamped Execution correction below the execution record if you find a mismatch.
State what actually happened in the correction.

The correction now preserves the actual sequence without replacing the plan. A run that followed the plan needs no correction.

State why the divergence happened in a second correction sentence.
Leave the original plan unchanged.
Leave the frozen list unchanged.
Leave every original timestamp unchanged.

What stays sealed?

The original plan records your intended method. The frozen list records what you remembered before the report influenced you.

An appended correction preserves what happened during execution. This keeps hindsight from rewriting the audit trail.

Confirm that the FROZEN timestamp predates the Header captured timestamp.
Confirm that the Header captured timestamp predates the transcribed findings.
Compare the recorded release wording with the publisher anchor.
Trace every number to the header, a report row, or a frozen-list item.
Confirm that report coverage matches the header exactly.
Confirm that the zero-setup cost path remains unchanged.

What should the integrity check show?

The frozen timestamp appears earlier than the header timestamp.
The header timestamp appears earlier than every copied finding.
The publisher anchor still names Claude Code v2.1.239 with the exact recorded wording.
Every count points to report evidence or sealed-list evidence.
Coverage never claims more sessions than the header states.
No installation, new account, metered API key, payment card, or added service entered the project path.

Stop the clock and write the verdict

The thirty-minute bar covers every action from Project opened through the written decision. Your verdict also needs a condition that could change it later.

Choose a practitioner verdict

Choose ADOPT when the evidence added enough specific value to justify the routine you named.
Choose REJECT when the report added no worthwhile evidence under a sample you judge sufficient for this trial.
Choose WATCH when a useful signal exists alongside unresolved uncertainty.
Choose PARK when the sample, timing, or prerequisites are too weak for a decision.

A missing total leaves the report without a denominator. That evidence gap supports WATCH or PARK.

Add the final evidence review to the worksheet by pasting this template:

Plain overview: I built one evidence-based evaluation of Claude Code /insights.
Artifact 1: Frozen pre-run record.
Artifact 2: Timestamped report record.
Artifact 3: Final evidence review.
Clock: [[PROJECT_OPENED="Project opened time"]] to [[WRITTEN_VERDICT="Written verdict time"]] = [[MEASURED_MINUTES="measured minutes to one decimal place"]] minutes. Thirty-minute bar: [[THIRTY_MINUTE_RESULT="met or not met"]].
Setup: Nothing installed.
Command path: The command shipped with the existing Claude Code client.
Browser path: The browser read the HTML file written by the client.
Platform: Run on Windows.
Equivalence status: [[EQUIVALENCE_STATUS="tested or not tested"]]. Untested equivalence remains reasoning.
Scope: Frozen pre-run record, report record, and final evidence review.
Dated anchor: Claude Code v2.1.239 was published at 2026-08-21T19:54:23Z.
Publisher wording: “Fixed `/insights` echoing literal `<analysis>` tags in its response on some models”
FROZEN at: [[FROZEN_TIME="FROZEN timestamp"]]
Header captured at: [[HEADER_CAPTURED_TIME="Header captured timestamp"]] before findings were read.
Four measurements:
1. Sessions analyzed and total: [[SESSION_COVERAGE="analyzed of total, or report absence wording"]]
2. Frictions confirmed: [[CONFIRMED_COUNT="number of confirmed frictions"]]
3. Frictions new: [[NEW_COUNT="number of new frictions"]]
4. Items on sealed list the report did not raise: [[NOT_RAISED_COUNT="number of sealed-list items not raised"]]
Evidence limit: Matching uses human judgment. No equality test determines it.
Coverage limit: The report covers recent eligible sessions on this machine only. Other devices and claude.ai are excluded.
Cost path: No metered API key or payment card was added. Existing plan usage was consumed.
Action limit: Reading suggestions leaves working habits unchanged until you act.
Verdict: [[VERDICT="ADOPT, REJECT, WATCH, or PARK"]]
Reason: [[VERDICT_REASON="evidence supporting the verdict"]]
Trigger that would reverse this verdict: [[REVERSAL_TRIGGER="specific reversal condition"]]
Re-check date: [[RECHECK_DATE="date when WATCH or PARK applies"]]

How does the final review work?

The opening fields preserve the project path and all three deliverables. The ordering fields prove that the prediction came before the report evidence.

The measurement fields carry your traceable counts into the decision. The final fields define the verdict as a revisable practitioner judgment.

Complete the review through the Action limit field using the evidence you already checked.
Complete the decision fields using the verdict conditions above.

You should now see a complete evidence review with only the final clock calculation left to finish.

Record the current Windows clock time to the second as Written verdict.
Subtract the Project opened time from the Written verdict time.
Round the elapsed minutes to one decimal place.
Complete the Clock line with the measured result.
Mark the thirty-minute bar as met or not met.
Enter Not required for this verdict for the re-check date when you choose ADOPT or REJECT.
Save the completed project worksheet.

Before your final read-through, do you expect every claim in the review to point back to a timestamp, report row, header value, or frozen-list item?

Read the final review aloud from the plain overview through the re-check date.

What should the final review contain?

All three deliverables appear by name.
All four measurements appear without a substituted summary.
The clock shows both timestamps plus elapsed minutes to one decimal place.
The platform remains limited to the Windows run.
The ordering evidence proves that the frozen record came before the findings.
The coverage limit excludes other devices plus claude.ai.
The cost path states that existing plan usage was consumed.
The verdict includes a reason, reversal trigger, and applicable re-check date.

🙋‍♀️ Final review has a gap?

A blank measurement means its source has not been traced back to the header, report rows, or frozen list.

A missing decision condition means the verdict cannot be revisited consistently later.

Help me audit my final evidence review.

You have turned an unsupported first impression into a dated decision you can defend. The frozen baseline remains ready for a future comparison.

Secret mission

Adjudicate the Closest Match

Your final counts depend on the judgment behind each match. Audit the least obvious classification by placing its exact evidence side by side without changing your sealed prediction.

Clean Up Your Resources

Your completed worksheet and local Claude Code report are stored on this Windows machine. Keeping them creates no ongoing cost.

Cost warning

This project adds no metered API key or payment card. Stored files do not consume plan usage.

Running /insights again consumes your existing plan allowance.

Resources you used:

The completed project worksheet containing the frozen record, report header, four measurements, verdict, and closest-match audit.
The generated /insights reports and cached analysis data in %USERPROFILE%\.claude\usage-data.

What stays outside cleanup?

Your local session history existed before this project. The project used it as a source without altering it.

The Claude Code client was also present before the project. Its installation remains untouched. Its sign-in remains untouched.

Keep everything running

No action is needed. Choose this option if you want to keep the report available for comparison with a future run.

Keep the completed project worksheet as your durable comparison baseline.
Leave %USERPROFILE%\.claude\usage-data in place so the browser report remains available.
Return on the re-check date in your worksheet if your verdict is Watch or Park.

Pause - I'll come back to this later

Shut down the open session to step away while keeping every artifact. Your existing local files create no ongoing cost.

End the existing Claude Code session.
Close the report tab in your default browser.
Keep the completed project worksheet in its existing saved location.
Return on the re-check date in your worksheet if your verdict is Watch or Park.

Your worksheet and report files stay on disk. Reopening them later requires no new installation or account.

Delete - I don't want to use this again

Remove the generated analysis data to start fresh. Delete the worksheet too only if you do not want the future comparison baseline.

This cleanup is broader than one report. It removes every /insights report plus all cached analysis data in the usage-data folder.

Your authentication remains untouched. Your preferences, installed plugins, and local session history remain untouched.

Delete the completed project worksheet from its existing saved location if you do not want the future comparison baseline.
Confirm the worksheet no longer appears in that saved location.
End the existing Claude Code session.
Delete the generated reports and cached analysis data by running this command in a Windows Command Prompt terminal:

rmdir /s /q "%USERPROFILE%\.claude\usage-data"

What does this command remove?

The command deletes %USERPROFILE%\.claude\usage-data plus everything stored beneath it. This includes report.html, timestamped report copies, and cached analysis data.

Return to the existing report tab in your default browser.
Refresh the report tab.

The local HTML report can no longer load from disk. Cleanup is complete because the generated report data is gone.

Still see the report?

Refresh the existing browser tab once more.
Check that the cleanup command targeted %USERPROFILE%\.claude\usage-data exactly.
Ask for help checking the Windows cleanup.

Nice Work!

You did it! You completed an evidence-based audit of Claude Code /insights. Your worksheet now supports a time-stamped routine decision with traceable evidence.

You've learned how to:

Create a frozen pre-run record before opening the report. Protect the sealed prediction from hindsight bias.
Preserve the count-bearing report header before reading findings. Use its timestamps to prove evidence ordering. Use its counts to state sample coverage.
Apply one-to-one classification to produce four traceable measurements. Turn those measurements into a time-counted Adopt, Reject, Watch, or Park verdict. State a reversal trigger. Limit the claim to this Windows machine's report sample.
Secret Mission: Audit the closest match by placing the exact frozen wording beside the exact report wording. Preserve any changed count or verdict as a visible correction. Keep the sealed list unchanged.

Ready to quiz yourself?