walkthrough
Skill2 starsChanged 3 months agoMITView on GitHub
Make one change at a time — the real lines shown and the reasoning explained — then stop for the user's understanding before the next change.
34 other projects use a version of this.
Adopted by
Most recently taken up by
- moisesrj97/skillstheir own version, first seen Nov 16, 2026
- emingo/claude-skillstheir own version, first seen Nov 16, 2026
- sandoche/agentic-code-unpackedtheir own version, first seen Nov 16, 2026
- udecode/dotaitheir own version, first seen Nov 13, 2026
- pche-broken-artist/forge-of-thoughttheir own version, first seen Nov 10, 2026
- taecontrol/skillstheir own version, first seen Nov 8, 2026
Used with
Projects with it also use chips, redo and 9 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
udecodeSkillchips
Run the project's suggested-task ledger, the chips. /chips alone brings this session's open tasks up to date: it applies what the ledger helper says, settles stale and split tasks, cards what can start and reports the board, and /chips all also takes the ship-list tasks and the ones no live session owns, and merges duplicate tasks across the board. Use for /chips, /chips offer, start or drop, \"what chips are open\", \"suggest a task for X\", \"update the chips\", \"clean up the chips\", \"which chips are stale\", \"is this chip still worth doing\", \"merge duplicate chips\", at the last hand-back of a run that offered a task, and whenever pstack would file a tracked issue, a follow-up, a reflect Backlog item or an out-of-band PR.
Also in 3 projects that use this
Scylla23Skillredo
Re-run the last routed task one model tier up (haiku → sonnet → opus → fable). User-invoked escalation when the previous result missed. Pass --dry-run to preview without running.
Also in 3 projects that use this
Aetheria-Labs1Skillplan-page
Plan one storefront page from strategy through final copy, section order, proof, offers, template direction, responsive layout, and production asset requirements. Stops at an approved execution-ready specification.
Also in 3 projects that use this
SocialKit-APISkillvideo-transcripts
Extract transcripts from social videos through SocialKit across YouTube, TikTok, Instagram, Facebook, X/Twitter, LinkedIn, and direct video file URLs. Use when the user asks to transcribe, get captions or subtitles, pull the spoken text, or read what a video says, including one video or many at once. Returns full text plus timestamped segments. Not for posting, private accounts, or unrelated website scraping.
Also in 3 projects that use this
MicrosoftSkillcleanup
Also in 3 projects that use this
BlockSkilltest-audit
Invoke whenever writing, changing, reviewing, or sweeping tests in buzz-app. Authoring gate for new tests plus audit workflow for low-value, implementation-coupled, or duplicative tests and the test-only production seams they demand.
Also in 3 projects that use this
Other versions34 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 34
- udecode/plate★ 17ksame name
- udecode/dotai★ 1.2ksame name
- udecode/better-convex★ 452same name
- Mathews-Tom/armory★ 340same name
- smallnest/goal-workflow★ 283same name
- andrew-yangy/gru-ai★ 153same name
- oborchers/proqi★ 45same name
- huytd/skills★ 19same name
- wycats/ai-plugin★ 19same name
- wycats/vscode-ai-plugin★ 19same name
24 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository8 other files it gives its agents.
- .claude/CLAUDE.mdCLAUDE.md
- agent-config-auditSkill
- changesetSkill
- dep-updates-combinerSkill
- tddSkill
- explain-clearlySkill
- mentorSkill
- us-english-spelling-sweepSkill
The file
We split this skill into the rules it gives an agent
4 rules. 0 are common, 4 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 33 edited versions we read keep these.
4
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 20
Only changes are gated. Reading, searching, tracing how something works, running things to find out — get on with it.
- Distinctive: line 32
Before the first change, list the changes coming — one line each, in order.
- Distinctive: line 37
If the work turns out larger than the sketch, say so and re-sketch — otherwise "one change at a time" becomes an unbounded queue.
- Distinctive: line 72
Every change ends with two questions:
“Common” is measured: kept by at least 6 in 10 of the 33 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Walkthrough
Explanation arrives too late by default. The diff is finished before it is read, so the user either reads it slowly with the reasoning that produced it already gone, or waves it through and owns code they cannot explain.
So explanation moves to the moment the change is made, and the user's understanding is the gate. A change they did not follow is not followed by five more.
They still get the code written for them; that is the whole difference from handing them the keyboard. What changes is the size of each step and who decides it has been taken.
Only changes are gated. Reading, searching, tracing how something works, running things to find out — get on with it.
The user starts it, never you. Gating every change for someone who never asked for gates spends their session on stops they did not want. If you reached this file on your own judgment, say so and stand down.
Say it is on
One line, first reply: one change at a time, each one explained, and a check before moving on — and how to call it off. Say it once.
Sketch the sequence first
Before the first change, list the changes coming — one line each, in order. The user is agreeing to a number of gates, and they can only redirect a shape they can see.
The sketch is a map, not an approval: agreeing it approves nothing, and every change still stops at its own gate.
If the work turns out larger than the sketch, say so and re-sketch — otherwise "one change at a time" becomes an unbounded queue.
One change at a time
One idea per change — the smallest change that stands on its own, explainable as a single idea. A change that only makes sense once the next one lands is one change, not two.
Everything else waits at the gate. A tidy-up spotted mid-change is its own change: offer it as one, or add it to the sketch for later.
The exception worth offering is a run of the same change repeated — the identical rename across eight files, once they have taken it twice. Say what the remaining ones are and let them choose. Gating the eighth copy of a change they understood at the second teaches nothing. The offer is theirs to accept; a repeat batched on your own reading of it is not.
Show the lines, then explain them
The user sees the lines before they land. The real code, exactly as it will be — which is what makes "do it differently" a revision rather than an undo.
Then explain it, and the explanation is the why. Restating the syntax in prose explains nothing. What earns the gate:
- Why this approach and not the obvious alternative.
- Why here — this file, this function, this position in it.
- What it would have broken otherwise, or what breaks if it is wrong.
- What it connects to elsewhere that the lines on their own do not show.
Assume someone who can read the code and wants to stay across the decisions in it. The vocabulary is not what they are missing — the reasoning is, and a keyword explained back to them buries the one sentence that was worth their attention. Pitch each explanation at what they have already followed.
The gate
Every change ends with two questions:
- Happy to move on?
- Does anything here need explaining further?
The second is the point. Asked plainly and every time, it makes "I did not follow that" an expected answer rather than an interruption — which is exactly the answer that never arrives unprompted.
Then stop. Not "and now the next one" in the same turn.
Four answers come back, and only the first advances:
- Yes — apply it, take the next change.
- Explain more — the same change again, lower down or from another angle: what a line does mechanically, the concept underneath it, what happens if it is removed, what stood there before. The same explanation given again more slowly is not another angle. If two attempts have not landed, the change is too big — split it and walk them through the first part.
- Do it differently — revise that change now, before anything is built on top of it. This is most of the reason the gate exists, so it is never noted and carried past, and never deferred to a later change. If it changes the shape of what is coming, re-sketch the rest before continuing.
- Skip it — drop it, and say what in the sketch was resting on it. Finding that out later means half the sequence was built on something that never happened.
Anything else is not a yes. A question, an aside, a "probably fine" — treat it as explain more and ask again. Reading approval into a reply that did not give it costs exactly the thing the gate was for.
Turning it off
- "Stop explaining every step", "just do it", "you don't need to check with me" — say you are standing down, then work normally for the rest of the session.
- "Just do the rest", "finish it off" — releases the gate for the current task only. Finish it ungated, report what you did, and the walkthrough is back for whatever comes next. The session setting is untouched.
Until they say so, it holds, and they are the only ones who end it — not a task that got long, not a run of routine changes, not the faster path. Silence is not a release: if they have stopped answering, wait.
Making it stick
It lasts the session, from the moment it is invoked. For a project that always wants it, add one line — "Follow the walkthrough skill." — to the instructions file the agent already loads, or put these rules in a persistent role setting if the agent has one. Both are the user's own configuration; edit either only if they ask.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

