PStack - Setup Guide¶

Source: backnotprop/pstack via skills.sh - 300.3K combined installs across 61 indexed listings; snapshot 2026-10-28 (Oct 28 morning sweep)

GitHub: backnotprop/pstack - 3,030 stars, MIT, branch main, plugin.json v0.15.34, 61 skill directories (58 under skills/, 3 under automations/benny/skills/), plus agents/poteto-agent.md and agents/comment-sicko.md

Category: Engineering discipline, review and verification workflow for coding agents

Quality Tier: 🟡 Beta - author-maintained mirror of cursor/plugins/pstack kept in sync through a two-branch flow, 3,030 GitHub stars, 61 indexed listings against 61 skill directories, three auditor verdicts sampled this sweep with one Snyk warning

pstack is Lauren Tan's (poteto) working skill set, published as a standalone mirror of cursor/plugins/pstack so it installs without Cursor. The MIRROR.md file explains the sync model: an upstream branch holds Cursor's files exactly, main is upstream plus this mirror's edits, and harness-neutral rewrites replace Cursor-specific behaviour so the skills run in any agent that reads SKILL.md files. The premise in the README is narrow and worth quoting because it explains the whole design: AI writes too much slop code, so the goal is not to maximize lines written, it is to write less but higher quality code, and throughput without quality is not a goal.

The suite is structured around one router. poteto-mode is the default entry point: it matches a request to one of 23 playbooks (investigation, bug fix, perf, hillclimb, runtime forensics, feature, refactoring, prototype, visual parity, eval, babysit, shipping, autonomous run, orchestrate, autopilot, session pickup, multi-phase plan and more), opens a todo list whose first items are that playbook's steps copied verbatim, and then routes to the other skills as the steps fire. Around it sit 24 short principle skills (one rule each, indexed inline so the router reads the index instead of every file), 33 task skills from unslop and interrogate to the verification family, and two subagent definitions: poteto-agent, the wrapper every delegate runs through, and comment-sicko, a read-only comment reviewer invoked through no-comments.

Installation¶

The mirror's own README opens with the skills.sh CLI, which installs from this repository into an agent's skills directory:

npx skills add backnotprop/pstack

The CLI lists every skill for search, then asks which skills and which agents. The upstream README documents the Cursor and Claude Code plugin routes instead:

/plugin marketplace add cursor/plugins
/plugin install pstack@cursor-plugins

Claude Code prefixes plugin commands and agents, so /poteto-mode becomes /pstack:poteto-mode, and the adapter in docs/harness/claude-code.md maps the remaining Cursor names. Claude Code leaves auto-update off for third-party marketplaces, so the install stays on the version you picked until you enable auto-update in the marketplaces tab or run claude plugin update pstack@cursor-plugins.

Hermes has no plugin marketplace, so the directory copy is the route that applies. Every skill is a plain skills/<name>/SKILL.md folder, which is the format the skills CLI installs anyway:

git clone --depth 1 https://github.com/backnotprop/pstack.git
cp -R pstack/skills/poteto-mode pstack/skills/unslop pstack/skills/interrogate \
  ~/.hermes/profiles/corpusiq/skills/

Several skills carry dependencies the README pins explicitly. teach needs how and why, architect needs arena and how, blast-radius needs arena, how, why and unslop, and the verification pair needs its maintainer (create-verification-skill with maintain-verification-skill). poteto-mode needs the principle skills and most of the rest, since it routes into all of them. The README calls out 17 skills that call nothing else and work alone, unslop, bro, how, tdd, typescript-best-practices, arena, swarm, interrogate, reflect, show-me-your-work, figure-it-out, automate-me, correct, benchmark-checklist, in-your-own-words, sharpen and statechart.

What It Provides¶

Skill	Installs	Use For
unslop	7,021	Cleaning up writing before it ships: removes the tells that mark text as machine written, across replies, docs and PR descriptions.
how	6,255	Walking through how a subsystem works, from the entry point to the data it touches, so a new contributor or an agent can act on it instead of guessing.
principle-fix-root-causes	6,139	Tracing each symptom back to the first place the system went wrong and fixing it there, so a guard is only accepted when the state it handles is legitimate.
why	6,059	Recovering why something was built this way: discovers the available MCP servers at run time, then queries source control, the issue tracker, long-form docs, chat, infra observability, error tracking and analytics in parallel.
interrogate	6,020	Putting a diff in front of several models and asking each to break it from an independent angle, including a strict code-quality lens.
technical-writing	5,943	Writing or revising docs, RFCs, READMEs, PR descriptions and commit messages against a layered standard built from Diataxis, Google developer style, Simplified Technical English and Global English.
arena	5,917	Getting N parallel attempts at the same task and taking the best parts of each, instead of committing to the first plausible answer.
poteto-mode	5,910	The rigour entry point: matches the request to one of 23 playbooks, writes the playbook's steps into a todo list, then routes each step to the skill that owns it.
architect	5,898	Settling the caller's usage, types and module shape before writing code that crosses a function boundary.
blast-radius	5,773	Answering what else a small looking change could break, and proving the safe fact by running code rather than asserting it.
reflect	5,770	Capturing a long task as a skill edit, so the next run of the same work starts from the recipe instead of rediscovering it.
figure-it-out	5,762	Designing a rigorous, auditable playbook for work that no bundled playbook fits, including long migrations the operator steps away from.
no-comments	5,736	Stripping comments before review, spawning the read-only Comment Sicko reviewer, fixing the findings it accepts and offering an encoding for each constraint it cannot verify.
show-me-your-work	5,733	Keeping a reviewable decision trail in a TSV an operator can read, commit or diff after an unattended run.
principle-type-system-discipline	5,732	Making illegal states unrepresentable: branded primitives, external data parsed at the boundary, exhaustive variants, types derived from the authoritative schema.
setup-pstack	5,723	Detecting the models you actually have access to and writing the rule that maps each role (code, judgment, the review panels) to one of them.
typescript-best-practices	5,720	Editing TypeScript with the type-system principle in view, expressed as syntax rather than abstraction.
create-verification-skill	5,705	Generating a project-local verify skill with a feature map when the project has no scripted way to prove app behavior, for any language or platform.

The remaining 43 indexed listings range from 9 to 5,689 installs, led by principle-prove-it-works (5,689) and maintain-verification-skill (5,664), then 21 more principle skills between 4,560 and 5,650, the review and process skills swarm (5,650), tdd (5,459), teach (5,443), correct (4,589) and benchmark-checklist (4,486), the utility set recall (5,581), make-bot-ui (5,283), poteto-help (4,145), ketchup (3,372), in-your-own-words (3,018) and sharpen (2,671), the two formal verification skills (maintain-formal-verification-skill at 1,605 and create-formal-verification-skill at 1,530), ketchup-video (827) and statechart (820), and the three benny automation skills at 9 installs each.

Why This Matters for Hermes Agents¶

The common failure in agent work is not a wrong answer, it is an unverified claim. An agent edits a file, reports it works, and nobody has run anything. principle-prove-it-works states the countermeasure directly: verify against the real artifact, run the feature, read the actual value, inspect the diff, and treat "it compiles" as a proxy rather than proof. That single rule changes what a Hermes session is allowed to call done.

The parallel story is the other half. poteto-mode exists so one agent can be trusted on deep work and several can then run at once, and the skills that make the fleet safe are the delegation principles: principle-guard-the-context-window routes bulk to subagents and keeps summaries in the main thread, while principle-never-block-on-the-human says proceed, present the result and reserve confirmation for irreversible actions. For cron and unattended runs that is the useful default, and it is stated as a principle rather than left to the operator's judgment.

The subagent wrapper is portable enough to matter here. poteto-agent must read poteto-mode in full, including its inline principles index, before doing any work, because substituting a general purpose subagent skips that read and drifts. In a harness with no subagent tool the skill says to do each delegated step yourself in order, which is exactly the shape of a small Hermes session.

Verification can also be generated rather than described. create-verification-skill builds a project-local verify skill with a feature map for any language or platform, maintain-verification-skill repairs that map when the app drifts from it, and the formal pair covers the cases tests cannot reach: races, retries, state machines, parsers and unsafe code. benchmark-checklist closes the loop by vetting a measured number (limiter, tuning, errors, repeat runs, end-to-end relevance) before anyone reports it.

Usage¶

You say	What happens
"this pr has a subtle bug where the scroll drifts every 750ms even when idle"	poteto-mode matches the bug fix playbook, writes the repro first, traces the drift to its cause and fixes it there, then verifies against the running app before reporting.
"why is this feature flag not on yet?"	why discovers the reachable MCP servers, queries source control, the issue tracker and chat in parallel, and returns the decision trail behind the flag rather than a guess.
"will this refactor break anything else?"	blast-radius maps the callers and dependents, then proves the safe fact by running code, and reports what it actually checked.
"review this diff like different models would, including quality"	interrogate sends the diff to several reviewers, one of them the strict code-quality lens, and returns the findings that survived the challenge.
"i'm going to bed, land the stack even if ci flakes"	poteto-mode takes the overnight playbook: independently verify the green stack, then land the contiguous verified run bottom-up through GitHub, without a question back unless an action is irreversible.
"this project has no way to prove the app works"	create-verification-skill generates a project-local verify skill with a feature map, then maintain-verification-skill keeps it honest when the app changes.
"set pstack up for the models I actually have"	setup-pstack detects the models on the machine, writes the role mapping to a rule every skill reads, and a rerun keeps only the roles you overrode.

Verification¶

Probe the paths you plan to install from. All six URLs below returned 200 when this guide was written, on branch main:

for p in skills/unslop/SKILL.md \
         skills/poteto-mode/SKILL.md \
         skills/interrogate/SKILL.md \
         skills/setup-pstack/SKILL.md \
         skills/create-verification-skill/SKILL.md \
         agents/poteto-agent.md; do
  curl -sL -o /dev/null -w "%{http_code} $p\n" \
    "https://raw.githubusercontent.com/backnotprop/pstack/main/$p"
done

Read three of them before installing:

https://raw.githubusercontent.com/backnotprop/pstack/main/skills/poteto-mode/SKILL.md
https://raw.githubusercontent.com/backnotprop/pstack/main/skills/no-comments/SKILL.md
https://raw.githubusercontent.com/backnotprop/pstack/main/skills/principle-never-block-on-the-human/SKILL.md

poteto-mode owns the playbooks and the principles index, no-comments carries one of the two community fixes this mirror adds on top of upstream, and the delegation principle governs what a skill will do without asking.

After install, check what landed. The principle set is the cheapest count to confirm if you installed poteto-mode with the skills it routes into, and the router's frontmatter should match the name the install index uses:

ls ~/.hermes/profiles/corpusiq/skills | grep -c '^principle-'
head -5 ~/.hermes/profiles/corpusiq/skills/poteto-mode/SKILL.md

The head output should show name: poteto-mode and disable-model-invocation: true, which means the skill is not auto-invoked from a description match. If the count is not 24, the principle set is incomplete and the router is reading an index it cannot follow.

Security¶

skills.sh verdicts for sampled skills (verified 2026-10-28) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
unslop	Pass	Pass	Pass
how	Pass	Pass	Pass
principle-fix-root-causes	Pass	Pass	Pass
why	Pass	Pass	Warn
interrogate	Pass	Pass	Pass
technical-writing	Pass	Pass	Pass

Five of the six sampled skills pass all three auditors. why carries a Snyk warning with code W011, which is published on that skill's security page, so read it before you install that one.

The install surface is markdown only. There is no package.json at the repository root, so adding these skills runs no lifecycle hooks. Two package.json files do exist inside skill directories, skills/poteto-mode/scripts/ (alongside bootstrap.ts, check-plan.mjs and the orch and watch-pr TypeScript entry points, with a bun.lock) and skills/statechart/scripts/. Nothing runs them at install time, but the router can call them during a run, so decide deliberately whether those two skills belong in a session that has network and repository write access.

The run-time surface is broader than the file surface. These skills drive git, gh, worktrees, subagent spawns and model routing, and the shipping and autonomous run playbooks take a verified stack through to merge rather than stopping at a proposal. setup-pstack writes a configuration rule that every other skill reads, so the model mapping it produces governs which model receives judgment calls. In Claude Code that file is ~/.claude/rules/pstack-models.md, written by /pstack:setup-pstack. Treat that rule as configuration with authority rather than as a preference, and review it once after the first setup run.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified 2026-10-28: 300,310 combined installs across 61 indexed listings; 3,030 GitHub stars; MIT; last pushed 2026-10-27. Counts drift over time.
The repository is a mirror, not the origin. Upstream is cursor/plugins/pstack, main is upstream plus local edits, and MIRROR.md warns that copying Cursor's files onto main erases those edits. Two community fixes are mirror-only, no-comments keeping a comment about an outside rule it cannot check and --base <branch> for orchestrate scripts on PR stacks that do not start at main, so a skill of the same name can behave differently here than in the Cursor plugin.
Three of the 61 indexed listings are benny automation skills under automations/benny/skills/, at 9 installs each. They are not registered as slash skills, they are Cursor automations only, and setup copies the pack into a target repository rather than installing it into an agent.
57 of the 58 directories under skills/ carry disable-model-invocation: true, so most of the suite will not fire from a description match; you name the skill you want, or you name poteto-mode and let it route. setup-pstack is the exception.
Some referenced material is not bundled. /deslop, control-cli and control-ui ship in the cursor-team-kit plugin and /create-skill is a Cursor built-in, so the full set needs that plugin alongside pstack.
The documented routes are skills.sh, Cursor and Claude Code. Hermes has no plugin marketplace, so a directory copy is the only path, and it brings no slash commands, subagent types or model routing with it.

Skills Catalog | Skills Marketplace

Pro Tips¶

Start with poteto-mode and let it route. The 23 playbooks are the actual product, and asking for one directly ("use the perf playbook") gets you the same steps without the router having to guess.
Run setup-pstack first thing. The default panel is opus 5.5 for judgment and grok for code delegation, and on any other model set the routing rule that gets written is the difference between a run that uses your models well and one that fights them.
Ask for the todo list before the work. poteto-mode copies the playbook's steps in verbatim as the first todo items, so the list is your review surface, and correcting it there is cheaper than correcting the patch.
Keep show-me-your-work on for unattended runs. The TSV decision trail is the only artifact that tells you why an agent chose a path once the session is gone.
When a number matters, pair benchmark-checklist with principle-explain-the-number. One vets the measurement, the other finds what limits it, and together they catch the speedup that measured something other than the change.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ