Harden install prompt against read-then-apply refusals (bolt.diy report)#63
Merged
Conversation
A real bolt.diy refusal showed that naming the checklist's steps isn't enough for agents that stage commands for the user instead of running them: "finish what its checklist flags" still reads as blind execution of unseen tool output. Reword the prompt tail to "review the checklist it prints, and make the two remaining changes yourself" — read-then- apply, with the agent as reviewer — identically in README.md, GETTING-STARTED.md, and field-test/prompt.txt. Encode the pressure as field-test/personas/bolt-diy.md and record it as refusal mode #5 in field-test/README.md. Gate on 2026-07-14: hostile 2/3 green (both refusals decisively cited the sub-hour v0.3.8 publish, i.e. documented release-freshness noise); green rounds credited the new wording as suspicion-lowering. Re-gate after 2026-07-16. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Prompt wording enhanced to require agent review before applying checklist changes. 🎯 Quality: 99% Elite · 📦 Size: Small 📈 This month: Your 220th PR — above team average · Averaging Excellent |
mariojgt
approved these changes
Jul 14, 2026
…t-diy-prompt-output Conflict resolution: main's bolt-diy persona wins (the verbatim bolt.diy system prompt with documented provenance) over this branch's hand-written pressure reconstruction, keeping the "personas embed the actual system prompts" table truthful; the usage comment is reworded to match. Flags line takes main's generic `--persona <name>` form. Refusal mode #5 and the retest-loop note from this branch are kept. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
A real-world bolt.diy refusal showed that naming the checklist's concrete steps isn't enough for agents that stage commands for the user rather than executing them — "finish what its checklist flags" still parses as blind execution of unseen tool output ("package trust ≠ output trust"). The prompt tail now reads "review the checklist it prints, and make the two remaining changes yourself", byte-identical across
README.md,GETTING-STARTED.md, andfield-test/prompt.txt. The pressure is encoded as a newbolt-diypersona and recorded as refusal mode #5 infield-test/README.md, which also folds the persona into the flags and retest loop.Gate
Ran 2026-07-14: hostile 3 rounds → 2 fully green (8/8); both refusals (hostile r3, bolt-diy) decisively cited v0.3.8 being published under an hour earlier — the documented release-freshness noise, not the new wording. Green rounds explicitly credited the new clause ("the CLI output never became the authority for what I changed"). Re-gate after 2026-07-16 per the plan-B ladder.
🤖 Generated with Claude Code