agentleFS
CsHengCsHeng/agent-skills/skills/implement-change/SKILL.md

implement-change

Skill3 starsChanged 8 days agoMITView on GitHub

Implement an authorized bounded change—including multi-file, multi-repository, or operational updates under settled contracts—or an approved plan through mutation, proportionate verification, conditional review, and bounded repair. Undecided goals or boundaries belong to design-change; unresolved execution ordering or coordination belongs to plan-change.

17 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use audit-test-suite, change-data-safely 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.
  • yaohuangguanSkill

    audit-test-suite

    Find where an existing test suite is weak and decide what to test next, ranked by risk. Use this when asked to add missing tests, find gaps, raise coverage, assess test quality or see which areas are thin, even if the request only mentions a coverage percentage.

    Also in 3 projects that use this

  • yaohuangguanSkill

    change-data-safely

    Change a database schema or stored data without breaking running code or losing data. Use this for any migration, new or changed column, backfill, bulk update, delete, index, or change to a stored format, including in small projects. Data mistakes are often impossible to undo.

    Also in 3 projects that use this

  • SauravSrivastavSkill

    debug-failure

    Diagnose a failing test, command, or runtime behavior using evidence and produce a minimal verified fix.

    Also in 3 projects that use this

  • MicrosoftSkill

    fix-flaky-test

    Reproduces and fixes flaky or quarantined tests. Tries local reproduction first (fast), then falls back to CI reproduce workflow (reproduce-flaky-tests.yml). Use this when asked to investigate, reproduce, debug, or fix a flaky test, a quarantined test, or an intermittently failing test.

    Also in 3 projects that use this

  • srclightSkill

    refactor-safely

    This skill should be used when the user asks to "refactor", "rename", "move", "change a function signature", "extract a method", "delete unused code", "clean up", or any modification that could break callers or dependents. Provides a pre-change safety checklist using srclight's graph tools to assess blast radius before touching code.

    Also in 5 projects that use this

  • +6 more from yaohuangguan

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 16 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
  1. Distinctive:

    When an action would exceed a real goal, hard constraint, commitment, risk, or permission, obtain the narrow decision for that impact first …

    line 22
  2. Distinctive:

    When existing edits overlap the intended write set, inspect their content and ownership, continue when they can be retained safely, and …

    line 43
  3. Distinctive:

    Prefer a narrow reproducer or red-green oracle for non-trivial behavior, with focused regression evidence for affected sibling paths.

    line 50
  4. Distinctive:

    When accepting a candidate or finishing a slice, keep each abstraction, guard, test, fixture, and document that names a behavior the …

    line 54

“Common” is measured: kept by at least 6 in 10 of the 16 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.

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

Implement Change

Complete the user's authorized outcome with the smallest durable change. The plan organizes the work; the user's goals, fixed choices, and real boundaries bind it. development-standards is the baseline for necessity and proportionate evidence.

Use This Skill When

  • an explicit bounded user request or approved plan authorizes repository mutation
  • implementation, verification, conditional review adjudication, and bounded repair need one semantic owner

A bounded request is sufficient on its own; it needs no separate design or plan artifact. Undecided goals, boundaries, or acceptance go to design-change; unresolved ordering, cross-actor coordination, or authority gates go to plan-change.

Preconditions

  • Confirm the authorized objective, approved order, repository ownership, allowed surfaces, protected behavior, acceptance evidence, recovery policy, and known user changes from the approved plan or, without one, from the request and applicable project contracts.
  • Inspect git status --short and preserve existing work.
  • Treat task membership, authority, and capability as separate facts. Commit, push, publication, deployment, destructive history changes, and external mutation each need matching authority beyond implementation approval. Consume existing trusted permission for the same target and side effects without re-asking; standing permission adds no work to a narrower request, and a design-only or source-only request keeps its scope until the user adds the delivery work. A push carries its entire outgoing history and the CI/CD it triggers.
  • When an action would exceed a real goal, hard constraint, commitment, risk, or permission, obtain the narrow decision for that impact first and continue independent authorized work meanwhile.

Execute The Settled Change

Implement the settled goals with the selected technology. Local investigation, glue, ordinary tests, same-task repair, and files the change causally requires are implementation work; an initial write inventory is context, not a whitelist. A needs-decision result describes an unresolved conflict for its owner; it does not reopen design or planning on its own.

Build a thin real path with useful local feedback first. For a replacement, follow references/replacement-and-cutover.md in development-standards: necessary prechecks, cutover, real-path verification, fix-forward repair. When supporting work grows without advancing useful behavior, reconsider the mechanism before assigning more workers or repairing every defect it creates; move non-prerequisite verification after cutover within existing authority. A known defect still gets a disposition against the actual outcome and risk.

Goals And Discretion

Main goals, necessary conditions, user-fixed technical choices, exclusions, explicit user-fixed checks, and inviolable boundaries bind. Plan approval binds required outcome and boundary acceptance at the delivery endpoint; model-inferred task granularity, dependencies, ordering, guards, compatibility assumptions, and local proof methods stay revisable within that intent, and are corrected without new approval when the required outcome and real boundaries hold. Authorized best-effort secondary work is adapted or omitted on evidence while the main goal continues, with the discretion and its effect recorded; a required result keeps its label and its oracle.

Investigate facts and generate execution products rather than asking for them. After the user answers a blocking question, resume the original authorized work. When a later user instruction changes task authority, reconcile it with the current objective and endpoint and, where the record update is authorized, rewrite the authorization block as one current account as described in plan-change's delivery reference.

Continue until the authorized outcome is delivered or a named boundary stops it; verification is complete when the evidence covers the changed behavior, affected consumers, and stated acceptance. Host interruption, user cancellation, and explicit invocation budgets end work as blocked with the remainder named.

During long tasks keep the objective, current result, remaining work, real blockers, and effective authority recoverable at milestones and handoffs. plan.md is an authorization, boundary, and handoff record; transient step tracking lives in working context or the host's passive task list, and plan.md changes only to reconcile an explicit authority, scope, or decision change or to record a handoff or terminal summary. Completion is the main agent's judgment from the actual result and evidence; task tracking adds no acceptance, test, or review obligation.

Checkout And Concurrent Work

  • Continue in the current checkout when uncommitted files are this task's own material or unrelated non-overlapping changes; preserve the latter and keep them out of this task's edits and any authorized commit.
  • When existing edits overlap the intended write set, inspect their content and ownership, continue when they can be retained safely, and otherwise resolve that specific conflict before writing. A reset, stash, or overwrite is a resolution only when its owner agrees.
  • Judge concurrency from the user's statement, host task evidence tied to the same checkout, or unexpected changes observed during this task, covering shared generated output and Git index or ref operations as well as file overlap. When drift appears, read and reconcile it, then continue under the existing goals and authority. Host exact-write and candidate-apply checks stay binding; refresh or redispatch through the supported parent path.
  • Read Worktree Isolation when this task needs its own Git worktree: an explicit request, an applicable repository rule, a required branch state, or a concrete conflict the current checkout cannot absorb. Host-managed isolation for delegated tasks needs no separate worktree procedure and grants no new Git mutation authority.

Implement And Verify

  1. Inspect the smallest current surface needed for the next ready task. For a bug fix in shared behavior, trace the symptom through the entry point, shared function, and affected callers to the contract that owns the invariant, using symbol, reference, configuration, and test evidence; choose a shared fix when the affected callers share the invariant and keep caller-specific rules at their owning boundary. When an unresolved caller contract prevents a safe bounded fix, state the gap and return the fitting outcome.
  2. Prefer a narrow reproducer or red-green oracle for non-trivial behavior, with focused regression evidence for affected sibling paths.
  3. Make the smallest cohesive change that advances required behavior within the approved scope: least total responsibility and maintenance burden, not fewest lines.
  4. Complete all approved in-scope tasks whose dependencies can be satisfied; a task boundary is progress, not a stopping point. Read Delegated Execution before choosing delegation slices and when dispatching or continuing a worker, interpreting a result, or recovering a failed invocation.
  5. Check changed surfaces against the user's goal, acceptance, authority, and fixed boundaries. Admit a discovered issue when the change caused it or it demonstrably blocks the authorized outcome; report unrelated existing failures without absorbing them. Reassess a growing mechanism before treating each resulting defect as a new goal.
  6. When accepting a candidate or finishing a slice, keep each abstraction, guard, test, fixture, and document that names a behavior the outcome depends on, and fold or delete the rest, including ones a plan or review named. A requested broader audit uses references/simplification.md in skills-routing.
  7. Run one verification strategy for the change. Required outcome acceptance, exclusions, and explicit user-fixed checks bind; choose the lowest-cost sufficient local proof for the actual diff from affected behavior, consumers, and uncertainty, and skip intermediate simulations a plan proposed when deterministic unit or schema checks plus authorized real-path verification already prove the boundary. Finish related edits before repeating checks that cannot inform intermediate work. The main agent selects and times integration, E2E, and full-suite checks at useful joins, uses a full suite for relevant broad impact, an explicit delivery requirement, or uncertainty narrower evidence cannot resolve, and reuses applicable worker results.
  8. Obtain independent review under the conditions review-change owns; when it runs, supply one bounded target with the objective, diff, and current evidence.
  9. Adjudicate findings and repair accepted defects with Focused Implementation Repair, rerunning the evidence the repair invalidated and retaining the rest.
  10. Complete when the authorized objective, delivery endpoint, necessary verification, and accepted findings are satisfied. Report the delivered result, its evidence, authorized secondary tradeoffs, and any unverified remainder; missing downstream permission or capability blocks that step only.

Delegation And Profiles

Optional execution and thinking profiles are provider-neutral hints; an absent or unmapped profile means a compatible default or keeping the task local. An explicit user choice of model or thinking level is preserved through ephemeral task parameters for that invocation; durable route configuration changes only under separate authority. Delegation transfers bounded investigation, implementation, verification execution, diagnosis, and authorized repair labor; scope control, continuation decisions, cross-task synthesis, finding adjudication, final acceptance, and the final response stay with the parent. When no independent runnable work remains while workers run, end the turn and act on the completion notification.

Recovery

  • fix-forward is the default: preserve evidence, diagnose the observed failure, repair within scope, and rerun the owning oracle.
  • stop-and-diagnose preserves current state and stops further mutation when continued mutation would compound uncertainty.
  • guarded-rollback runs only under an approved exact trigger, target, and rollback verification, with evidence that rollback is safer than forward repair; repeated failure is not that evidence and does not reopen design.

Outcomes

  • ready: implementation and required verification pass and no accepted finding remains. Disclosed authorized secondary tradeoffs may remain when the main goal and approved acceptance hold; a requested production endpoint that was not performed is reported as not performed.
  • needs-decision: qualified as plan when the approved scope, order, or verification is insufficient, design when evidence invalidates an approved design boundary, or authority when completion needs new user or external authority.
  • blocked: required evidence, a prerequisite, or capability is unavailable, diagnosis yields no new evidence and no in-scope path remains, or work stopped on cancellation or an explicit budget. Name the actual reason, what was tried, and the remaining work.

Report what changed, the evidence, and material limits or unfinished work; include review or repair conclusions when they affect acceptance, and report only actions actually performed.

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.