agentleFS
wheeloswheelos/apollo-lite/.agents/skills/long-horizon-development/SKILL.md

long-horizon-development

Skill177 starsChanged 7 days agoApache-2.0View on GitHub

Deliver long-running, multi-stage features and integrations by staying anchored to the user goal, completing a minimal working path, establishing a baseline, then filling gaps and optimizing. Use when implementation spans multiple milestones or can stall on unavailable data or metrics.

The file

Show its rules and text

We split this skill into the rules it gives an agent

14 rules. About 0 are common, 14 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈14

    Distinctive · estimate

    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
  1. Distinctive:

    Use for substantial implementation that spans multiple milestones, modules, or dependencies and needs continued delivery without losing the …

    line 10
  2. Distinctive:

    Do not use this skill for a small, bounded change, a root-cause investigation, or a narrowly scoped algorithm, performance, build, or …

    line 14
  3. Distinctive:

    Read the repository's agent rules and the relevant module contracts, source, tests, and worktree status.

    line 21
  4. Distinctive:

    Do not turn open evidence into a prerequisite by default.

    line 42
  5. Distinctive:

    Prefer existing module boundaries, interfaces, and test mechanisms.

    line 53
  6. Distinctive:

    Keep unavailable data behind an explicit input boundary.

    line 54
  7. Distinctive:

    Do not use empty stubs, fabricated success, silent fallbacks, or mock output as evidence of production behavior.

    line 58
  8. Distinctive:

    After the minimal path works, record its observable behavior and verification evidence: relevant tests, outputs, configuration, input …

    line 65
  9. Distinctive:

    Do not treat a functional baseline, fixture result, or proxy as a performance or real-world effectiveness baseline.

    line 74
  10. Distinctive:

    After each increment, check whether it still advances the stated outcome.

    line 83
  11. Distinctive:

    Do not treat checkpoint status as authorization or a test result.

    line 113
  12. Distinctive:

    Follow-on increments are scoped to the goal, and deferred items have a reason and a condition for revisiting them.

    line 121
  13. Distinctive:

    Stop only the affected path when a hard gate fails, required interfaces conflict, or missing evidence makes a correctness or safety claim …

    line 127
  14. Distinctive:

    If blocked, report the specific decision or evidence needed, what can proceed, and the consequence of deferral.

    line 132

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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

Long-Horizon Development

Scope

Use for substantial implementation that spans multiple milestones, modules, or dependencies and needs continued delivery without losing the user's goal. The task must include implementation, not only research or design.

Do not use this skill for a small, bounded change, a root-cause investigation, or a narrowly scoped algorithm, performance, build, or regression task. Route those to the corresponding skill. This skill coordinates delivery; it does not replace domain-specific contracts or validation.

Prerequisites

  • Read the repository's agent rules and the relevant module contracts, source, tests, and worktree status.
  • Confirm the user-visible outcome, existing baseline, constraints, and boundaries. Keep established work and module ownership intact.
  • Identify what is already known and what data, metrics, or decisions are unavailable. Ask only when the missing information changes safety, correctness, a hard contract, or the implementation boundary.

Workflow

1. Keep the goal and blockers explicit

State the intended outcome in one sentence and identify the smallest observable result that would demonstrate progress toward it. Separate:

  • Hard gates: Safety, correctness, required interfaces, legal constraints, and user-confirmed release or acceptance criteria.
  • Open evidence: Data, measurements, or metrics that would improve confidence but are not required to build and verify the initial path.

Do not turn open evidence into a prerequisite by default. Do not invent a target metric or claim evidence that has not been collected. If missing evidence is a hard gate for a production claim or behavior, keep that boundary explicit while continuing safe, independent work.

2. Deliver the minimal working path first

Choose the smallest vertical slice that exercises the real ownership and interfaces from input to an observable result. Implement the framework and minimum behavior needed for that path before broadening scope.

  • Prefer existing module boundaries, interfaces, and test mechanisms.
  • Keep unavailable data behind an explicit input boundary. Use a fixture or synthetic input only to verify the contract or plumbing, and label what it does not validate.
  • Represent unsupported, unavailable, or diagnostic-only behavior explicitly. Do not use empty stubs, fabricated success, silent fallbacks, or mock output as evidence of production behavior.
  • Validate the slice with the smallest repository-supported check that proves its stated behavior.

3. Establish and preserve a baseline

After the minimal path works, record its observable behavior and verification evidence: relevant tests, outputs, configuration, input source, and known limitations. Use real representative data when available; otherwise record the fixture or synthetic input and its limited scope. Missing operational data or product metrics must not stop work that can be implemented and checked without them.

For a performance change, use performance-optimization to capture the relevant pre-optimization measurements and compare like-for-like results. Do not treat a functional baseline, fixture result, or proxy as a performance or real-world effectiveness baseline.

4. Complete in goal-aligned increments

Turn the remaining work into small increments with observable outcomes and clear acceptance checks. Prioritize by user value, dependency, and risk; keep each increment independently reviewable and verifiable.

After each increment, check whether it still advances the stated outcome. Defer polish, broad data collection, speculative abstractions, and non-blocking metrics when they do not change correctness or the next decision. Revisit deferred items when evidence becomes available or a hard gate depends on them. Do not require a plan document, fixed milestone count, or agent delegation unless the task or repository requires it.

5. Complete and optimize from evidence

Fill required behavior and integration gaps before expanding into optional capabilities. Add representative data, metric coverage, and scenario validation as they become available or necessary for a stated acceptance claim.

Optimize only against a real requirement and an appropriate baseline. For algorithm design or implementation, hand off the relevant work to algorithm-development; for measured resource bottlenecks, use performance-optimization; for behavior requiring replay or closed-loop evidence, use simulation-regression. Keep the delivery goal, interfaces, and deferred risks coherent across these handoffs.

Acceptance

For work that must survive interruption, use the checkpoint procedure. Store task state in the session workspace, not in durable context. Capture every relevant checkout, untracked input, configuration, and fixture in the explicit fingerprint scope. On resume, compare fingerprints, recheck current ownership and outstanding approvals, then reuse only evidence whose inputs are unchanged. Do not treat checkpoint status as authorization or a test result.

  • The user-visible goal and hard gates remain explicit throughout delivery.
  • A working, inspectable minimum path exists before optional breadth or polish.
  • Baselines and evidence identify their inputs and limitations; no proxy is presented as real-data, production, or performance validation.
  • Data or metrics that are not hard gates do not block independent, verifiable implementation.
  • Follow-on increments are scoped to the goal, and deferred items have a reason and a condition for revisiting them.
  • Completion and optimization claims match the evidence actually obtained.

Failure handling

Stop only the affected path when a hard gate fails, required interfaces conflict, or missing evidence makes a correctness or safety claim impossible. Preserve completed work and continue independent, safe increments where possible.

If blocked, report the specific decision or evidence needed, what can proceed, and the consequence of deferral. Do not silently weaken requirements, invent data or metrics, broaden scope, or present a scaffold as a finished feature.

Method reference

Adapt the small, independently testable task boundaries, explicit interfaces, and checkable outcomes from Superpowers writing-plans (reviewed 2026-10-26). This skill deliberately does not require a long implementation plan, test-first workflow, commits, or subagents for every task.

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.