agentleFS
Ex-imanityEx-imanity/SkillCollections/skills/deep-task-preflight/SKILL.md

deep-task-preflight

Skill1 starsChanged 17 days agoMITView on GitHub

Use before work when the goal, scope, completion criteria, or evidence boundary is unclear; when a previous attempt was rejected for missing context or incomplete verification; when the user explicitly asks for a plan, acceptance criteria, comparison with existing implementation, or a deeper preflight; or when a user-facing, cross-system, irreversible, or multi-phase deliverable needs alignment before editing. Build a compact evidence-based preflight, stop only for blocking questions, and verify the final result against the acceptance matrix. Do not trigger for routine one-line edits, direct factual answers, or standalone verification requests unless the user also asks for planning or alignment.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Do not promote a task to Full merely because it has several steps.

    line 39
  2. Very specific:

    Before planning edits, read applicable AGENTS.md, README, local skill instructions, branch/worktree state, and the closest existing …

    line 44
  3. Distinctive:

    Use the full or compressed template in packet-template.md.

    line 54
  4. Very specific:

    If implementation is authorized and the gate is READY, show the packet and continue without ceremonial approval.

    line 89
  5. Distinctive:

    Keep every change tied to an acceptance ID and inspect the diff/artifact after each meaningful phase.

    line 95
  6. Distinctive:

    Automated check: run the cheapest meaningful lint, unit test, schema check, render check, or deterministic comparison.

    line 98
  7. Distinctive:

    Consumer read-back: read through the same path that consumes the write or artifact (page, report, cache, downstream service, export, or …

    line 104
  8. Very specific:

    Do not report “done” because files were created or tests merely started.

    line 110
  9. Distinctive:

    Read only the relevant bundled reference before execution.

    line 116
  10. Very specific:

    Use repository-code-style for local implementation conventions, context-resilient-task for durable recovery, and cross-agent-review only …

    line 130

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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

Deep Task Preflight

Purpose

Turn “think more deeply” into a reviewable workflow: align the intended outcome, boundaries, evidence, decisions, and completion checks before changing files or external state. Depth comes from coverage and falsifiable checks. Never request or reveal hidden chain-of-thought; record concise conclusions, sources, assumptions, alternatives, and decisions.

This skill supplies the alignment gate. Read only the relevant adapter in references/adapters/ for domain-specific verification. Follow repository instructions and domain skills in addition to this workflow.

Choose a mode

  • Lightweight: one clear, reversible change with an obvious check. State the outcome, one acceptance check, the scope boundary, and the next action.
  • Medium: a few files or one subsystem, with a known goal but non-trivial compatibility or validation. Use the compressed packet template; keep criterion IDs, unknowns, the gate, and the validation matrix.
  • Full: unclear or previously disputed work, cross-repository/system changes, audience- facing deliverables, irreversible or data-affecting work, or tasks that need multiple sessions/agents. Use the full packet and relevant adapters.

Do not promote a task to Full merely because it has several steps. Promote it when a wrong interpretation would change scope, behavior, evidence, or completion.

Baseline and reference-first rule

Before planning edits, read applicable AGENTS.md, README, local skill instructions, branch/worktree state, and the closest existing implementation or artifact. When an existing client, server path, parser, renderer, schema, workflow, or template is named or plausibly relevant, compare it on the same representative input before designing a replacement. Record the reference path/revision, input and output shapes, transformation rules, and intentional differences. If no reference exists, say so. A similar-looking file is not evidence of equivalent behavior.

Preflight packet

Use the full or compressed template in packet-template.md. Full and Medium packets must retain observable outcome, scope boundary, criterion IDs, unknowns, the gate, and a validation matrix. Full mode additionally records audience, source priority, change surface, references, risks, and alternatives.

The outcome must describe the user's result, not the agent's activity. “Add three slides” describes work; “help the audience understand the problem, motivation, evidence, and result within three slides” describes an outcome.

Challenge checks before editing

In Medium and Full modes, read challenge-checks.md and apply its nine checks: stranger, traceability, boundary, evidence, negative space, failure path, rework, reference parity, and end to end. Record failures in the packet and revise it before editing; do not compensate with a longer task list.

Ambiguity, stopping, and re-gating

Classify each uncertainty:

  • Blocking: different answers change the goal, scope, public behavior, ownership, safety boundary, or acceptance result. Set NEEDS INPUT, ask at most three questions in one batch, and do not mutate the workspace.
  • Non-blocking: state the assumption and proceed using local conventions or stronger evidence.
  • Testable unknown: perform the cheapest read-only inspection or validation that can resolve it.

NEEDS INPUT exits when the user answers the questions; then rebuild the affected packet sections and re-run the gate. If the user does not answer, leave the workspace unchanged and report exactly what is waiting. BLOCKED requires a concrete missing dependency, access failure, contradictory source, or unsafe boundary plus an exit condition; do not claim partial completion as done. Re-run the preflight gate whenever new evidence changes the goal, an acceptance criterion, the change surface, a referenced contract, or the risk boundary.

If implementation is authorized and the gate is READY, show the packet and continue without ceremonial approval. If the task expands beyond the packet, pause at the next safe boundary, update the packet, and gate again.

Execution and universal verification ladder

Keep every change tied to an acceptance ID and inspect the diff/artifact after each meaningful phase. For every changed path, use this input → transform → output loop:

  1. Automated check: run the cheapest meaningful lint, unit test, schema check, render check, or deterministic comparison.
  2. Boundary check: inspect the actual request, state, artifact, or payload at each changed interface; verify identifiers, counts, ordering, grouping, empty values, references, and freshness. Do not assume the first observed item is complete when the contract permits multiple outputs.
  3. Consumer read-back: read through the same path that consumes the write or artifact (page, report, cache, downstream service, export, or device) and confirm the observable state changed. A success response or HTTP 200 is not completion evidence.
  4. Acceptance closure: revisit every acceptance row, attach exact commands/results and evidence paths, label partial or unverified criteria, and state remaining risks.

Do not report “done” because files were created or tests merely started. For cross-session or multi-agent work, persist the packet and decisions with context-resilient-task when a .task-state exists or local conventions require durable recovery.

Route to an adapter

Read only the relevant bundled reference before execution. The adapter describes representative inputs, boundary checks, consumer read-back, failure cases, and evidence to retain. If no adapter fits, use the universal verification ladder and state the task-specific boundary checks in the preflight packet.

Task shapeReference
Code, refactor, compatibility, testscode-changes.md
Documents, presentations, reportsdocuments.md
UI and interactive stateui-interactive.md
Data extraction, batch, media, LLM/report pipelinesdata-pipelines.md
Debugging and incident analysisdebugging.md
API/schema/migration/deploy/security/performance workcontracts-and-operations.md

Use repository-code-style for local implementation conventions, context-resilient-task for durable recovery, and cross-agent-review only when an independent reviewer is requested or the risk justifies it. These skills complement the gate; they do not replace it.

Lightweight response

Render in the user's language:

Outcome: <observable result>
Acceptance: <one concrete check>
Scope: <one sentence including the main non-goal if useful>
Next: <immediate action>

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.