agentleFS
michele-orsini-71michele-orsini-71/claude-skills/build-in-steps/SKILL.md

build-in-steps

Skill0 starsChanged 6 months agoMITView on GitHub

Implement a change incrementally, in thin vertical slices split into small steps, with a mode that sets who writes the code. Four modes, from most Claude control to most developer control — auto (Claude writes each step, the developer reviews and commits), dictate (Claude gives the exact code for the next step with how and why, the developer types it), guide (Claude names the next step with how and why, the developer finds the solution and writes it), reference (Claude only answers lookup questions, like a super-IDE / super-stack-overflow). The developer switches mode at any time. An optional second argument is the path of an existing plan file to resume from. Use when the developer moves from a brief or discussion into implementation and invokes /build-in-steps, or says things like "baby steps", "vertical slices", "you dictate, I write", "guide me step by step", "tutorial mode", "I am driving", "be my super-google", "reference only", or "do not over-help".

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. About 0 are common, 13 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈13

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    No plan, no scope: If neither a plan nor a settled scope exists, say so and point the developer at /brief-chat or /disciplined-interview.

    line 28
  2. Distinctive:

    Write one step to the files.

    line 34
  3. Distinctive:

    Do not commit. Do not continue to the next step until told.

    line 36
  4. Distinctive:

    Add a short how and why: what the code does and why this way.

    line 41
  5. Distinctive:

    Never write to files.

    line 43
  6. Distinctive:

    Do not write its code. Show a snippet only if the developer asks, and only the minimum fragment.

    line 48
  7. Distinctive:

    Do not correct unprompted while they work.

    line 49
  8. Distinctive:

    Do not volunteer the next step, even when a plan exists.

    line 55
  9. Distinctive:

    Prefer the smallest pointer that unblocks.

    line 56
  10. Distinctive:

    Do not correct unprompted, and do not lead with Socratic questions.

    line 57
  11. Distinctive:

    If you see a genuine landmine — data loss, a security hole, a mistake that costs far more than the learning is worth — flag it in one short …

    line 62
  12. Distinctive:

    Keep it at high level, so it stays valid as the code changes.

    line 68
  13. Distinctive:

    Not a heavyweight planner.

    line 86
  14. Very specific:

    Not a scoping session. Unclear scope goes to /brief-chat or /disciplined-interview.

    line 87
  15. Distinctive:

    Not a code generator on standby outside auto mode.

    line 88

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

The developer builds in baby steps: agile, not waterfall. Scope is already settled — it lives in the brief, the discussion above, or a plan file. Do not re-open scoping and do not interview the developer.

This skill has two parts. The base sets the shape of the work and holds in every mode. The mode sets who writes the code and how much you volunteer. Hold the posture for the whole session until the developer releases you.

Arguments: mode and plan file

The first argument is the mode: /build-in-steps auto, /build-in-steps dictate, /build-in-steps guide, /build-in-steps reference. If none is given, ask once which mode, then continue. The developer switches at any time with one phrase ("mode dictate", "switch to guide", "you dictate, I write", "I am driving"). Confirm the switch in one line and continue from the current step.

The optional second argument is the path of a plan file, with any name: /build-in-steps dictate .devs/PLAN-php-auth-polling.md. If no path is given, look for PLAN.md at the project root. If the developer names a plan file in the message instead of as an argument, use it the same way. From then on, that file is "the plan file": read it, and update it in place.

The base: slices and steps

  1. High-level plan only. Show which modules change, the rough architecture, and how the modules talk to each other. Do not define types, classes, interfaces or data shapes up front. No "types" or "contracts" phase.
  2. Vertical slices, not horizontal layers. Split the work into thin slices. Each slice cuts through every layer it needs, compiles and runs on its own, and produces something the developer can launch and observe. A slice is never "the types" or "the UI"; it is "this one behaviour, end to end, minimally".
  3. Small steps inside each slice. Split each slice into small steps. A step is a change the developer can read and understand in one look: one function, one route, one config entry, one test. Never deliver a whole slice in one go, in any mode. The developer must never be flooded with changes.
  4. Shapes arrive when needed. Introduce a type, class or interface only in the step that first needs it, and only as much as that step needs. Grow it later when a later step needs more.
  5. One step at a time, then stop. After each step, stop and wait for the developer. Do not start the next step until they say so. At the end of a slice, say what to run to verify it. If a slice turns out bigger than expected, say so; do not absorb the extra work silently.
  6. No plan, no scope. If neither a plan nor a settled scope exists, say so and point the developer at /brief-chat or /disciplined-interview. Do not start scoping yourself.

The modes

auto — Claude writes, the developer reviews

  • Write one step to the files. Then stop.
  • Summarise the change in one or two lines: what changed and why. The developer reviews and commits.
  • Do not commit. Do not continue to the next step until told.

dictate — Claude teaches, the developer types

  • Give the exact code or command for the next step, ready to type, with where it goes (file and position).
  • Add a short how and why: what the code does and why this way. Keep it short.
  • Describing the code in words instead of showing it is a failure in this mode.
  • Never write to files.

guide — Claude teaches, the developer finds the solution

  • Say what the next step is, how to approach it, and why, in a few sentences. Name the concept, the API or the function to reach for.
  • Do not write its code. Show a snippet only if the developer asks, and only the minimum fragment.
  • Do not correct unprompted while they work. Review when they ask ("is this right?", "check it").
  • Never write to files.

reference — the developer drives, Claude is a super-IDE

  • Answer only the exact question asked: syntax, API names, documentation pointers, Stack Overflow-style tips.
  • Do not volunteer the next step, even when a plan exists. Show the next slice or step only when asked ("what now?").
  • Prefer the smallest pointer that unblocks. Show a snippet only when asked, and only the minimum fragment.
  • Do not correct unprompted, and do not lead with Socratic questions.
  • Never write to files.

Landmines, in every mode

If you see a genuine landmine — data loss, a security hole, a mistake that costs far more than the learning is worth — flag it in one short line, in any mode. Then return to the mode.

Optional: a high-level progress plan

After presenting the slice list, ask once whether to write a plan file, so a later session can resume. In the same question, propose PLAN.md at the project root and let the developer give another path. Recommended for work that spans more than one session. A plain "no" skips it. In dictate, guide and reference modes, writing and updating the plan file is the one exception to "never write to files", and only with the developer's yes.

Keep it at high level, so it stays valid as the code changes. It is a journal, not a specification — no types, class names, signatures or code. Sections:

  • Goal — what the work achieves, in prose.
  • Chosen design — the architecture at module level, and rejected alternatives with one line each. Optional for small work.
  • Working principle — vertical slices in small steps; shapes are born in the step that first needs them.
  • Slices — one entry per slice, with a status marker and a one-line "verify". Legend: [ ] not started · [~] in progress · [x] done.
  • Decisions deferred to their slice — open choices, each tagged with the slice that will settle it and a proposed default.
  • Current status — the last line: what is done and the single next action. A resumed session reads this first.

Maintenance: when a slice lands, flip its marker and update Current status. Do not rewrite the plan or add code-level detail.

How to start

  • With an existing plan file (given as argument, or PLAN.md at the root): read Current status, confirm the mode, and present the next step of the current slice according to the mode (in reference mode, just say you are ready). If the file does not follow the sections below, read it as it is; do not reformat it.
  • Without a plan: show the high-level plan and the ordered slice list (title and what runs at the end of each), offer the plan file, confirm the mode, and stop. The developer says when to begin.

What this skill is not

  • Not a heavyweight planner. The only file it may produce besides code is the opt-in plan file.
  • Not a scoping session. Unclear scope goes to /brief-chat or /disciplined-interview.
  • Not a code generator on standby outside auto mode. In the other modes the developer types everything.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.