Skillpack · v0.11.1
Fifteen skills I wrote for Claude Code and Codex. Most of them started as a prompt I kept typing by hand.
Every time I opened a fresh session, I asked the agent to devour the codebase before it touched anything. Every few weeks I asked it to check that the foundations still held. Every time I went deep on a concept, I ended up saying "totality" and sending it back for another pass.
Those prompts worked, but only when I remembered to type them. So I started packaging them. Each one became a skill: a written method an agent loads on demand, with its own rules, its own output, and a guard against its own worst failure.
The first was memory-scriber, in March 2026, because every new conversation started from zero. The pack got its name a month later. Most of what followed came the same way, from a prompt I kept retyping or a word I kept using.
“The point is not to collect prompts. The point is to package durable ways of working.”
Six stages. Every skill does one job in one of them, and the two sibling packs slot into the same loop.
Claude Code and Codex do the work, often side by side in herdr.
↺ anything consequential still unknown goes back to Understand
Skills stay single-purpose. They combine through three files at the root of a repo, and each one is read as a claim to re-check, never as settled truth.
The thinking skills leave nothing behind but the conversation:
Adding to a codebase
devourpotentialdecide ⇄ probebuildverify
From an idea to a project
exploreisomorph / scourdecidevisionfirst build
Hardening
devour refreshbedrockfixverify
Intent
Keeps what the project is for written down.
Three moments: birth, backfill and refresh. Evidence drafts it, I decide it, and the original spark stays verbatim. Drift gets raised with me, never resolved quietly.
LeavesVISION.md
“Evidence proposes, the visionary disposes.”
Understand
Studies a codebase before anyone changes it.
Maps structure, runtime routes, timing and blast radius (if you touch X, inspect Y), keeps an honest list of unknowns, and has to answer eight questions or declare the repo partially devoured. Collab mode sends a Claude and a Codex to map it separately, then makes them argue.
LeavesMAP.md
“My pain problem is every time I start fresh with an agent, I asked it to 'DEVOUR' the codebase.”
A reality check in the middle of a conversation.
Pulls out the claims a conclusion rests on and tries to break each one against fetched sources: confirmed, wrong, oversimplified, outdated or no consensus. It never answers from memory.
LeavesConversation only
“Can you scour the internet to make sure we're at the right page? That's the pain point.”
Maps a whole surface with coverage you can check.
Anchors against official taxonomies, walks the surface from several directions, expands every example into its full family, and keeps a shelf for the niche. It ends with the tree, an audit line and what is left over.
LeavesA coverage tree, or a report
“I want the totality of fruits, like 'oranges', and you gave me orange pictures.”
Deep research that ends in two files.
Covers every source it can reach, then converges into notes.md and one self-contained report.html instead of a pile of scratch files. Scan mode answers inline when a full report is too much.
Leavesnotes.md + report.html
Decide
Thinks about a system through a mature twin.
Finds a domain with the same structure that already paid for its mistakes, turns its laws into invariants and scores the system against them. A metaphor that produces no invariants counts as a failed run.
LeavesOptionally, the twin in VISION.md
“Isomorph isn't a bug finder. I would say it's more of an architect.”
Reads what the codebase wants to become.
Walks the vision, the map and the audit, then looks for capacity already there: squeeze, birth, combine, expose, generalize. No idea survives without structural evidence, someone who benefits, and a smaller alternative.
LeavesConversation only
“A pretty idea with no beams under it is decoration.”
Turns options into a direction.
Names the decision and who owns it, compares two to four options including doing less, attacks the favourite, then recommends one with its accepted downside and a signal for revisiting it.
LeavesConversation only· suggested by Codex in a review of the pack
Runs the smallest experiment that settles a question.
Declares the assumption, the test and the cleanup before running anything, works in isolation, and hands decide a verdict: supported, contradicted or inconclusive.
LeavesConversation only· suggested by Codex in a review of the pack
Build
Turns a vague feeling into a brief another agent can run.
Recovers the real goal, sorts what is given from what is assumed, applies a lens for the domain, and writes a brief with scope, authority and how to verify. Its main guard is against ornate wrongness.
LeavesA copyable prompt
“Max means maximum useful fidelity, not maximum length.”
Picks the right installed skill for the job.
Reads my skill catalogue and ranks two to five fits, with when to reach for each. Read-only.
LeavesConversation only
Verify
Audits the foundations like a building inspector.
Re-runs every kept repro first, then attacks: can it half-complete, fire twice, race, swallow a failure? Findings are confirmed or suspected and ranked silent-wrong first. Grades are earned only from logged attacks, and nothing gets fixed during the audit.
LeavesAUDIT.md + runnable repros
“You know how a bank is reliable? I call it bank logic.”
Plays the user and the machinery at the same time.
Drives a real journey end to end in an isolated, reversible copy, ranks every step by risk, then stress-tests the risky ones with varied inputs at volume and checks the results against an independent oracle.
LeavesA run diary + findings in AUDIT.md
“You're like 'user' and a 'god' of the app.”
Keep
Closes a session into memory, the way a colleague would.
Writes the opening brief verbatim, a reflective middle and the journey into the host's memory folder, then recalls it at the start of the next session.
LeavesA memory entry
“It's literally a human memory that I tried to mimic after having a meeting or talking to a colleague.”
Turns a workflow that worked into a new skill.
Needs two or more very different worked examples, picks a procedure shape or a thinking-mode shape, and writes a quality gate that can actually fail. isomorph, gauntlet and totality were all made with it.
LeavesA new SKILL.md
Two sibling packs sit in the same loop: hats for the roles an agent holds during live work, extra for keeping a record of it.
The closeout review before a diff lands.
A structured Codex review I run after build and verify.
LeavesReview findings· from OpenClaw's agent-skills
I fixed a bug in it upstream ↗Watches a live session while I drive.
I click through the app; it checks the screen against the API, the database and the logs, and records findings in my words.
LeavesFindings
Rebuilds my work trail from conversation history.
Day, agenda and domain notes reconstructed from my messages across many sessions.
LeavesAn Obsidian-style trail
A contribution record with receipts, plus a review essay.
Withdraws from the memory and the trail to reconstruct what someone did over an era and how they worked.
LeavesA record and an essay
The Fantopy record came from it →Claude Code
/plugin marketplace add github.com/kennykankush/skillpack
/plugin install workbench@kennykankush-skillpack
Codex
codex plugin marketplace add kennykankush/skillpack