agentleFS
doodledooddoodledood/manifest-dev/.claude/skills/harden-task-file/SKILL.md

harden-task-file

Skill78 starsChanged 5 months agoMITView on GitHub

Harden a manifest-dev task guidance file for one-shot quality — either /define''s quality-gate/Default set or figure-out''s probe set. Iterates: orthogonality gap analysis, user-approved additions, prompt review, fix, converge. Use when a task file needs comprehensive coverage or "harden task file".

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 4 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    If no arguments, ask which task file to harden.

    line 13
  2. Very specific:

    Write findings to /tmp/harden-{timestamp}.md after each round.

    line 34
  3. Distinctive:

    If the first analysis finds no gaps, invoke the reviewer once and exit if clean — not every task file needs hardening.

    line 57
  4. Distinctive:

    Log updated: after each round

    line 67
  5. Distinctive:

    Add arbitrary numerical thresholds

    line 101

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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

User request: $ARGUMENTS

Systematically harden a /define task guidance file until manifests built from it produce deliverables that don't need iteration.

If no arguments, ask which task file to harden.

Context

Task files come in two parallel, decoupled sets, each keyed by task type and hardened for one-shot quality. Harden whichever set the target file belongs to:

  • /define's task files (claude-plugins/manifest-dev/skills/define/tasks/) carry encoder data:
    • Quality Gates — Verifiable output properties. Can split into baselines (always enforced) and selectable (meaningful rigor choices)
    • Defaults — process practices always worth doing, carried into the manifest without probing; /define decides the encoding, and routes one whose violation would be unsafe or irreversible to a binding invariant rather than advice
  • figure-out's task files (claude-plugins/manifest-dev/skills/figure-out/tasks/) carry probing fuel:
    • Blind-spot probes — Non-natural angles the model skips by default (failure modes, pre-mortem fuel), phrased as the question that opens a branch
    • Forced trade-offs — Competing tensions the model must drive to a decision

New items must match the depth and structural conventions of the owning skill (define/SKILL.md or figure-out/SKILL.md) and its existing sibling task files.

Goal

The task file should be comprehensive enough that a manifest built with it surfaces all criteria needed for one-shot quality — figure-out's probes surface the non-natural angles during understanding; /define's gates and Defaults encode the verifiable bar. "One-shot" = the deliverable passes review without iteration.

Log

Write findings to /tmp/harden-{timestamp}.md after each round. Read full log before each new round — prevents re-proposing rejected items and losing dimension context.

Per-round log structure:

## Round N
### Dimension Map
[dimension → items mapping]
### Gaps Found
[uncovered dimensions]
### Proposals
[item: accepted/rejected by user]
### Reviewer Findings
[finding: agree/disagree, applied/skipped]

Orthogonality Analysis

The core discipline. Map every item in the file to a dimension — an independent axis of concern. A dimension is a top-level concern like "evidence quality" or "audience fit"; items are specific checks within a dimension like "source credibility" or "cross-referencing". Two items share a dimension if improving one naturally helps the other.

User validates the dimension map before gap-filling begins. Gaps = dimensions with no coverage.

Examples of dimension sources: deliverable lifecycle (creation → review → use → maintenance), rejection triggers, wrongness vs incompleteness, base-rate failures for this task class, user interaction points.

If the first analysis finds no gaps, invoke the reviewer once and exit if clean — not every task file needs hardening.

Iteration Loop

Each iteration achieves:

  • Gaps identified via orthogonality analysis
  • Additions designed — invoke the prompt-engineering skill before proposing changes
  • User-approved additions applied (all additions via AskUserQuestion)
  • Quality validated — invoke the review-prompt skill on the task file after applying changes
  • Reviewer findings evaluated critically — not all are valid. Present assessment with rationale; user decides
  • Log updated after each round

Converged when criteria in Convergence section met.

Section Placement

Each item belongs in exactly one section, and a safety rule whose violation would be unsafe or irreversible is a Default even when verifiable — that clause wins over the Quality Gate (baseline) row below, which every safety rule also matches. A file holds only its own set's sections — a /define file never carries probes/trade-offs, a figure-out file never carries gates/Defaults:

SectionSetWhat it checksTest
Quality Gate (baseline)/defineOutput property that should always be trueWould omitting this ever be acceptable? No → baseline, unless violating it would be unsafe or irreversible → Default
Quality Gate (selectable)/defineOutput property representing a meaningful rigor choiceReasonable to skip for some tasks? Yes → selectable
Default/defineProcess practice always worth doingCan't verify from output, or a safety rule whose violation would be unsafe or irreversible? → Default
Blind-spot probefigure-outA non-natural failure mode / pre-mortem angle the model skips by default"Imagine this was rejected because..." AND a capable model wouldn't raise it unprompted? → probe
Forced trade-offfigure-outCompeting tension with no universal right answerBoth sides have legitimate merit? → trade-off

A concern appears once, in its most natural section. When the same concern appears in multiple sections, keep the stronger version.

Principles

PrincipleEnforcement
Orthogonality over volumeCover all dimensions, not all possible items within a dimension
User approves all changesPropose via AskUserQuestion, never auto-add
Critical reviewer evaluationEvaluate each finding independently — push back with rationale when wrong. When reviewer suggests items in already-covered dimensions, orthogonality wins
No redundancy across sectionsSame concern in both risks and scenarios = pick one
Principles over thresholds"Corroborated across independent sources" not "verified across 2+ sources"
No capability instructionsDon't prescribe verification methods — parent skill handles that
Match complexity to domainNot every task file needs 17 quality gates — match depth to the diversity of ways the deliverable can fail

Never

  • Auto-add items without user approval
  • Blindly apply all reviewer suggestions
  • Add arbitrary numerical thresholds
  • Prescribe verification methods (parent skill handles this)
  • Re-propose items the user already rejected (check log)

Convergence

Done when:

  • Orthogonality analysis finds no new uncovered dimensions
  • Prompt reviewer finds no MEDIUM+ issues
  • User confirms satisfaction

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.