Agent Skills: demi

demi

A demi-glace starts concentrated. Nothing extra goes in just because the kitchen has it nearby. You do the same with code: build the smallest useful version that satisfies the ticket, fits the repo, and can be verified. Complexity has to earn its way into the pan.

Core principle: start simple enough that reduce should not be needed immediately afterward. Small means understandable and verifiable, not cramped. The shortest version that hides intent is not demi; it is just dense.

The demi ladder

Before writing custom code, stop at the first rung that holds:

1. Does this need to exist? If the current behavior already satisfies the actual ask, skip the change and say why.
2. Can the repo already do it? Use existing commands, helpers, components, config, conventions, and tests before inventing new ones.
3. Does the standard library cover it? Prefer built-in language or framework features over custom code.
4. Does the platform cover it? Native browser, database, shell, operating system, framework, or deployment features beat owned code when they fit.
5. Does an already-installed dependency cover it? Use what is already in the project. Do not add a dependency for a few clear lines.
6. Can one explicit local change do it? One function, one component, one endpoint, one flag, one small data path.
7. Only then: write the minimum custom implementation that works and can be checked.

The ladder is quick. Do not turn it into a research project. If two rungs work, take the higher one unless it is less correct on edge cases.

The demi pass

Run this before editing:

1. Restate the actual ask. One sentence. If the task is vague, choose the smallest useful interpretation that still satisfies it.
2. Find the local pattern. Read the nearest existing code, helpers, tests, and naming conventions. Prefer what the repo already uses.
3. Climb the ladder. Check existing behavior, repo primitives, standard library, platform, installed dependencies, and local one-change solutions before new infrastructure.
4. Name the smallest useful slice. Define the first complete behavior the user can actually run, click, test, or review.
5. Cut speculative scope. Remove features, settings, abstractions, options, data models, and UI states that are not required for that slice.
6. Pick the boring implementation. Use direct code, existing helpers, and local composition. New abstractions need evidence.
7. Set a growth trigger. Write down what future fact would justify expanding the design: a second consumer, a third repeated pattern, a measured performance limit, a public API boundary, or an explicit user requirement.
8. Verify the slice. Run the narrowest meaningful check through check. If behavior is hard to pin down, use taste for a focused test before expanding.

Simplicity gates

Before adding a new layer, answer these. If any answer is no, keep it local.

• Is this required by the current request?
• Does the repo already have a pattern for this?
• Is there a standard library, platform, or existing dependency path?
• Will at least two real callers use this now? Three is better for shared abstractions.
• Would a future maintainer understand this faster than the direct version?
• Can I verify the behavior after this change?

What to build first

Prefer these moves:

• One vertical slice over a broad scaffold.
• One explicit function over a generic utility.
• Existing components, commands, config, and test harnesses over new infrastructure.
• Clear local logic over premature indirection.
• A small typed data shape over a general schema system.
• A direct UI state path over a reusable state machine.
• A narrow CLI flag over an interactive prompt or new config file.
• A plain file or local module over a package, service, queue, plugin, or database.
• A small runnable check over a full testing framework when the repo does not already have one.

What to refuse early

Do not build these unless the current task proves they are needed:

• Framework swaps.
• Plugin systems.
• Generic registries.
• New persistence layers.
• Background workers.
• Multi-provider abstractions.
• Config formats for a single value.
• Theme systems for one screen.
• Large component libraries for one view.
• Repository-wide rewrites.
• Public API reshapes.
• "Future-proof" abstractions without a future requirement.

Not skimping

demi is not permission to be careless. Never simplify away:

• Trust-boundary validation.
• Error handling that prevents data loss.
• Security controls.
• Accessibility basics.
• Compatibility shims that protect real users.
• Calibration or tuning points for real hardware or external systems.
• Tests or checks for non-trivial branches, loops, parsers, money paths, auth paths, or security-sensitive logic.
• Anything the user explicitly asked for after you challenged the scope once.

A simple implementation without a check is not finished. Use the smallest meaningful check, not necessarily the biggest test harness.

Shortcut notes

If the small version has a known ceiling, name it in the plan or final report. Add an inline comment only when future maintainers would otherwise mistake the simplification for ignorance.

Examples:

• Global lock is fine for single-user CLI; move to per-account locks if concurrent accounts matter.
• Linear scan is fine below the current file sizes; index when measured input size makes it hot.
• Native date input is enough here; custom picker only if design or browser support requirements change.

The one-screen plan

For non-trivial work, write this before editing:

## demi: <task>

Actual ask: <one sentence>
Smallest useful slice: <what will work when done>
Highest rung that holds: existing behavior | repo primitive | stdlib | platform | installed dependency | local change | custom code
Existing pattern to follow: <file/helper/test/component>
Cut from scope: <what is intentionally not being built>
Growth trigger: <what would justify expanding later>
Verification: <command/check/user-visible proof>

Keep it short. This is a guardrail, not a design doc.

Mode

demi is apply-by-default. After the quick pass, implement the smallest slice unless the task is architectural, public-interface-heavy, security-sensitive, under-specified in a way that changes the outcome, or the user explicitly asks for options first. In those cases, report the demi plan and wait.

For large work, combine with recipe: use demi to keep each recipe task small, vertical, and verifiable.

Boundaries

• Use mise when context is missing.
• Use recipe when the work needs sequencing.
• Use taste when the slice needs tests before confidence.
• Use check before claiming done.
• Use reduce only after code exists and needs behavior-preserving cleanup.

Common mistakes

| Mistake | Reality | |---|---| | Building a full framework because the feature might grow | Growth is a trigger, not a prediction. Ship the slice. | | Adding config before there are real variants | One value can be code. Two may still be code. Three starts a conversation. | | Creating a generic utility from one use | That is not reuse, it is disguise. Keep it local. | | Scaffolding every future state | Build the path the user asked for and leave clean edges. | | Calling fewer files "less complex" | A thousand-line god file is not demi. Small means understandable, not cramped. | | Picking the shortest algorithm when it drops edge cases | Lazy means less owned code, not flimsier behavior. Correctness wins ties. | | Skipping local pattern discovery | Simple in isolation can be weird in the repo. Match the kitchen. | | Treating YAGNI as technical debt | YAGNI keeps quality high and scope small. Skimping leaves a mess for later. |