AgentStack
Browse Sign in
Browse Why AgentStack Sell Docs
Sign in
SKILL verified MIT Self-run

Prevent

skill-dtiger1889-ops-claude-harness-toolbox-prevent · by dtiger1889-ops

You just did something documented knowledge says not to do -- a rule, prior

No reviews yet
0 installs
0 views
view→install

Install

$ agentstack add skill-dtiger1889-ops-claude-harness-toolbox-prevent

✓ scanned · ✓ verified, works with Claude Code, Cursor, and more.

Security review

✓ Passed

No issues found. Passed automated security review. · v0.1.0 How review works →

  • Prompt-injection patterns
  • Secret / credential exfiltration
  • Dangerous shell & filesystem operations
  • Untrusted network calls
  • Known-malicious package signatures

What it can access

  • Network access No
  • Filesystem access No
  • Shell / process execution No
  • Environment & secrets No
  • Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

View the full security report →

Verified badge

Passed review? Show it. Paste this badge into your README, it links to the public security report.

AgentStack Verified badge Links to your public security report.
[![AgentStack Verified](https://agentstack.voostack.com/badges/verified.svg)](https://agentstack.voostack.com/security/report/skill-dtiger1889-ops-claude-harness-toolbox-prevent)

Reliability & compatibility

Security review passed
0 installs to date
no reviews yet
3d ago

Declared compatibility

Claude CodeClaude Desktop

Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.

Preview Execution monitoring

We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.

How agent discovery & health will work →
Are you the author of Prevent? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

prevent -- the mistake was preventable; make it impossible to repeat

The premise is FIXED: you did something the user and a past session already established you shouldn't do, OR something a documented rule forbids. The knowledge exists somewhere; you didn't hit it. This is a HARNESS BUG, not a one-off slip -- "I'll remember next time" is the failure mode this skill exists to kill. The deliverable is a durable change to the files future sessions read, so the same mistake becomes structurally hard to make.

Step 1 -- fix the immediate mistake

Same discipline as /dumb: name the wrong action in one line, no defending, then actually correct it now (undo the bad instruction, redo the operation the right way, retract the wrong advice). Don't let the harness work below stall the fix.

Step 2 -- find the knowledge that should have stopped you

Search for where this was already established. Grep across the workspace (if your workspace .gitignore hides project files from the ripgrep-backed Grep tool, use plain grep -rin), check the relevant project's CHECKPOINT.md and CLAUDE.md, archived specs, and past session transcripts if needed. Quote the documented fact verbatim with its file path. If it genuinely was never written down, say so -- Step 4 then writes it for the first time instead of moving it.

Step 3 -- diagnose WHY it wasn't read (pick one, say it plainly)

  • Wrong location -- the fact lives in a file this workflow never opens (e.g. a

deployment rule sitting in one project's CHECKPOINT, while the editing session only ever opens the deployed artifact's folder). Most common.

  • Buried -- the file WAS read, but the fact was one line in a wall of other state.
  • Wrong layer -- a stable rule stored as mutable state (CHECKPOINT) instead of in

CLAUDE.md, or a mechanical rule stored as prose instead of a hook/deny entry.

  • Not surfaced at decision time -- the rule is fine but the action happened on a

pre-oriented prompt / subagent / automated path that skips the file.

  • Never documented -- first occurrence was talked about but never landed in a file.

The diagnosis determines the fix; don't skip it and just paste the fact somewhere.

Step 4 -- move the knowledge into the path of the work

Place (or duplicate-with-pointer) the fact where the FAILING workflow cannot miss it. Choose the strongest applicable rung:

  1. Mechanical enforcement -- if the mistake is a command pattern, a hook or

settings.json deny rule beats any prose.

  1. Colocated doc -- a CLAUDE.md (or top-of-file comment / README line) sitting IN

the directory where the work happens, stating the rule in its first lines. Example: a deployed-artifact folder whose colocated CLAUDE.md opens with "edit the deployed copy and toggle the extension -- NEVER reinstall the installer package."

  1. Project CLAUDE.md (stable rule) or CHECKPOINT.md open thread / Key decision

(mutable state) of the project the workflow starts from -- respecting the 30-line and 120-line caps.

If your setup treats the harness files as the ONLY durable memory (this harness does -- see harness_me.txt on auto-memory), a fix that lands only in an auto-memory file is a FAILED /prevent: put it in the files future sessions actually read. If you do use auto-memory as your durable store, it counts as a rung -- but pick one store and commit.

If the fact already lives somewhere legitimate, don't delete it -- leave the detailed version where it is and put the short load-bearing line + pointer in the new location. Follow the harness conventions while doing this: move files with mv/Move-Item (not Write+delete), version-stamp bumps + changelog entries when a CLAUDE.md changes, line caps.

Step 5 -- report, short

Three lines: (1) what you did wrong and the fix applied, (2) where the knowledge was hiding and why it missed you, (3) where it now lives and why that location is in the read path of the next session that attempts this action. No promises about future behavior -- the file IS the promise.

Rules

  • A spoken "noted, won't do it again" without a file edit is a FAILED invocation of this

skill. Something durable must change on disk every time /prevent fires.

  • Don't over-correct into bloat: one tight rule line in the right place beats a paragraph

in three places. Respect the CLAUDE.md 30-line cap; spill detail into a pointed-to file.

  • If the right placement is genuinely ambiguous (two plausible homes), pick the one

closest to where the bad action physically happens and note the alternative in one line -- don't ask unless both placements would require restructuring.

  • If the same rule has now been violated twice from DIFFERENT entry points, that's the

signal to escalate a rung (prose -> colocated doc -> hook/deny).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet, be the first.

Versions

  • v0.1.0 Imported from the upstream source.