oscar/dotfiles
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(claude/skills/audit): forbid previewing upcoming tiers

Browse Source 
The model was appending a one-line heads-up describing the next
tier's findings, reading 'do not render a later tier' narrowly as
only barring the full finding block. Spell out that previews,
teasers, and counts of what follows are not allowed either.
This commit is contained in:
oscar
2026-07-04 17:24:12 +02:00
parent 6d52068755
commit c086e2b769
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
claude/skills/audit/SKILL.md
+2 -2
View File
@@ -56,7 +56,7 @@ The parent does not review code directly and does not read in-scope source files
4. Merge the reviewers' findings and drop exact duplicates. Do not present anything yet.
5. Validate the merged findings. For a small set (a handful), validate inline: read the referenced locations and nearby context and confirm each finding is real. For a large set (roughly eight or more, which deep runs usually hit), spawn a single validation subagent per `### Validator` instead, to keep the reads off the main thread.
6. Classify the confirmed findings into the tiers in `## Presenting and applying`. Classification only. Do not edit any file, including `TODO.md`, until the user approves the tier that contains it.
7. Present and apply one tier at a time. Show a tier, stop, and wait for the user before showing the next. Do not render later tiers in the same message.
7. Present and apply one tier at a time. Show a tier, stop, and wait for the user before showing the next. Do not render later tiers in the same message, and do not preview, tease, or summarize what is coming in later tiers.
The scope block passed to each reviewer:
@@ -199,7 +199,7 @@ Misfires are noisy. Validation keeps signal high.
## Presenting and applying
Findings go to the user one tier at a time, not as one big report, so they track a single decision at a time: present a tier, get approval, apply, commit (deep mode only), gate, then move to the next. Show one tier per message, stop, and wait. Do not render a later tier in the same message, even when you already know its findings. Skip empty tiers. If nothing remains after validation, output exactly `Nothing flagged` and stop.
Findings go to the user one tier at a time, not as one big report, so they track a single decision at a time: present a tier, get approval, apply, commit (deep mode only), gate, then move to the next. Show one tier per message, stop, and wait. Do not render a later tier in the same message, even when you already know its findings. Do not preview, tease, or summarize upcoming tiers, not even a one-line heads-up or a count of what follows. The current tier is all the user sees until they resolve it. Skip empty tiers. If nothing remains after validation, output exactly `Nothing flagged` and stop.
### Tiers