mcpField
C
computa-please

computa-please

github.com/khoahyhDevelopment0100 installsUpdated today
Repository

About computa-please

Computa Please is an agent skill for carrying software tasks from a concrete outcome through authorized implementation and verification. It emphasizes using existing working paths, preserving scope and authority, and stopping when proof or permission is missing.

Key features

  1. 01Start from concrete outcomes
  2. 02Inspect existing inputs and evidence
  3. 03Preserve scope and authorization
  4. 04Route work to relevant skills
  5. 05Verify proportionately and report blockers

Use cases

  1. 01Implementing a clearly specified change
  2. 02Choosing tools for external-service tasks
  3. 03Reviewing task authority before edits
  4. 04Verifying changes and reporting blockers

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/khoahyh/skills --skill computa-please

After installing, type/computa-pleasein Claude Code to confirm it loaded.

Bundled filesreferences/review.md

Details

From the project README

Computa Please

Start with the next concrete useful outcome and closest existing working path, data, and connection. Inspect source inputs and captures before arranging fresh access or tooling.

Scope

Keep the requested outcome, explicit decisions, exclusions, proof, and authority in working context. Discussion and planning are read-only unless edits are authorized. A clear implementation request needs no separate interview or spec.

Resolve remaining facts from repository instructions, canonical code, existing proof, and linked context. Choose routine details; ask through the question tool for decisions affecting behavior, scope, consequential risk, or authority, recommending an answer while continuing independent work.

Size complexity for the actual workload and concrete failure modes; synthetic valid bounds are boundary evidence, not expected scale. For each non-obvious addition, ask: what concrete failure of the agreed outcome would removing this cause? Preserve evidenced caller, data, security, and rollback obligations at their owning boundary. Surface necessary scope expansion before implementing it; splitting extra work into PRs does not justify it.

Tools And Access

Prefer an existing authorized structured tool for external-service work. When the capability or connection is unknown, check Executor when available before choosing computer use, arranging new access, or declaring access missing. Read Executor's own execute guide via skills({ name: "execute" }); search by service and operation, inspect matching schemas, and consult saved connections when needed. Reuse a known suitable path without repeating discovery.

For example, use Executor's connected Tailscale integration for tailnet administration when it supports the requested operation; configuring or troubleshooting the local VPN client is a separate task.

Use computer use for UI-only operations, requested UI verification, or an evidenced gap in structured tooling. If Executor is absent or lacks the operation, use another authorized existing path. Preserve the intended account and scope across fallbacks; report permission failures rather than routing around them. A connected tool establishes capability, not authority to change state.

Routing

Load only the guidance that changes this task's approach. Let the selected skill or repository contract own its procedure and output format.

RequestRoute
Unsettled product behaviorfeature-grill for decisions blocking this delivery; a single decision needs only a direct question.
Building, polishing, or reviewing web UI and interactionsemil-design-eng.
Mobile web or PWA touch, viewport, scroll, or safe-area behaviormobile-native.
React Native or Expo motion, gestures, transitions, or hapticsanimate-expo.
Finding places that would benefit from animationfind-animation-opportunities for read-only recommendations.
Naming a described web motion effectanimation-vocabulary.
Bug or performance regressiondiagnosing-bugs; verify the original failing path after an authorized fix.
OptimizationSet a baseline, objective, constraints, and stopping condition; measure the actual result.
Simplificationsubtract, or scope-prune for implementation drift.
New or changed teststest-audit's authoring bar and the target repository's checks.
PR-bound work or requested reviewReview for reviewer selection and reuse.
Existing review feedbackreview-remediation for the selected feedback set.
PR bodypr-body against the final diff.
Stacked PRs or dependent branch workgh-stack for GitHub's native Stacks.

For UI work, load only the matching guidance and preserve the project's design system and existing dependencies.

Use Git for repository state and commits, and gh for ordinary PRs. Keep focused work in one PR. Split independent changes into separate PRs against the repository's default branch.

Authority

Before writes, establish the task-owned checkout using worktrees and the repository's rules. Preserve unrelated work and stage only task-owned changes.

Standing grant (Khoa, 2026-10-06): in gocustodia/platform, gocustodia/infrastructure, and gocustodia/trowel, an implementation or fix request includes scoped additive commits on the task branch (not main), push, and opening or updating one draft PR after applicable local verification, including repair of task-caused required checks. Identify the repository by its remote; the grant covers its task worktrees.

Carry same-task grants through fixes and resumes. Other repositories need an existing or task-specific grant. Honor narrower task and repository restrictions; name the conflicting instruction when it blocks delivery. Readiness, review requests, merge, deployment, policy bypasses, destructive changes, history rewrites, unrelated work, and external messages need their own authority. Follow the project's operations contract for production effects, including merge-triggered deployments.

New PRs start as drafts. Establish affected branches, worktree owners, and command effects before stack mutations; selecting Stacks grants no stack-wide publication or history-rewrite permission.

Verify And Finish

Choose one risk-matched verification plan: smallest behavioral proof plus owning required gates, run once on the stable candidate. Reuse unchanged proof; rerun only evidence invalidated by material changes. CI may supply wider proof where repository policy allows; optional comfort checks are not gates.

Diagnose only failures blocking delivery; reuse a supported working fixture or path where sufficient. Retain the causal diagnostic; after two consecutive attempts yield no new evidence, switch paths or report the blocker. Resume diagnosis only with new evidence.

For changed UI, inspect the rendered result and exercise the affected interaction. Distinguish browser or emulation evidence from checks requiring real hardware.

Inspect the complete final diff against the agreed outcome, then finish through the authorized delivery boundary. Before merging, establish effects and authority; verify the merge and applicable post-merge results against the landed commit. Reconcile uncertain external actions before retrying.

Once proven, finish rather than extending review, polling, or cleanup cycles. If required proof, facts, or authority are missing, report the concrete blocker with the outcome and evidence. A pending action is not completed delivery.

Coordination

Use delegation, extra review, or artifacts only to resolve a specific uncertainty blocking delivery or satisfy a request or requirement. Limit concurrent subagents to two unless requested otherwise. Keep coupled work and shared resources under one owner; give delegates scope, path, authority, and required evidence, then integrate and verify consequential claims.

Persist only for requested artifacts, recovery, or coordination. Keep one current, compact handoff with the outcome, decisions, authority, state, evidence pointers, blockers, and next action. On pickup, reconcile it with live checkout and provider state. Keep secrets, customer data, and raw transcripts out of shared artifacts.

Comments

Markdown supported