Repository navigation
Releases
align 0.11.0
Prepare reviewed team decisions for an MLOps-owned training pilot. This release adds validated manual submissions and selected-target exports without collecting records or training a model automatically.
- Validate reviewer policy, approved records, artifact hashes, context, criteria versions, withdrawals and task-family splits.
- Export prompt/completion SFT, eligible text DPO, code evidence and selected tool actions. Keep the legacy text file for compatibility.
- Check local tokenizer boundaries and refuse silent truncation. Detect contradictory preference cycles.
- Read the MLOps brief or Korean version.
Validation: 67 repository tests passed. Three invented examples loaded in Hugging Face Datasets and passed a pinned Qwen tokenizer boundary check. These are compatibility checks, not model training or proof of improved coding performance. Reproduction artifacts and limits are linked in the brief.
Existing installations need the skill/plugin update through their supported route. Team export uses version-2 records and templates; personal version-1 text records remain supported. No agreements or collection settings are changed automatically.
ALIGNS.md v0.10.0: Optional training data and quieter reminders
Add optional path 8 to collect and review decision examples for user-run personal fine-tuning. Choose sources, scope, local storage and review checkpoints before collection. Export reviewed text-only SFT, eligible same-context DPO pairs and held-out evaluation data, with a separate provenance manifest. Training, Git commits and uploads remain separate choices.
Return invitations now appear at most once when an actual self-alignment session closes. Ordinary project tasks, installation-only work, intermediate items and skill development no longer append them. Existing managed instruction blocks need an authorized refresh; updating the skill alone does not replace those blocks.
English, Korean and Japanese READMEs are synchronized. The current package also includes the maintained method changes since earlier release archives.
Validation: 44 Python tests, the Skill Creator validator, and a CLI export smoke check passed. These checks do not establish model-training improvements or full cross-client conversational behavior. Tool-call trajectory export is not supported in this version.
ALIGNS.md v0.5.2: Review loop and goal mismatches
Review loop and goal settings, stopping decisions and completion claims when they miss the user's intent. The method compares the request, configured objective, run state and actual results, then clarifies only unresolved criteria.
Known errors, usage limits and requested pauses do not need a new alignment interview. Reviewing a mismatch does not authorize restarting, increasing the budget or overriding a user stop. The skill needs an active agent turn; it cannot wake a stopped process.
Validation includes 114 fresh Codex/Astra executions across English, Korean and Japanese. In the final 24-case loop/goal set, all nine alignment requests loaded the method and none of the fifteen controls did. The original 42-case discovery suite was also replayed. A broader description that caused unnecessary loads remains in the published evidence and was not released.
Requests, results and limitations
The ZIP contains the complete align skill and matches the tested final source snapshots. SHA256SUMS.txt verifies the archive. Updating the method does not change personal ALIGNS.md agreements.
ALIGNS.md v0.5.1: Validate natural-language discovery
A vague request to learn preferences now prompts for a concrete criterion before creating an agreement. Unrequested drafts stay in the conversation; explicit lasting preferences can still be saved directly. The skill name and discovery description are unchanged.
Validation used 60 synthetic English, Korean and Japanese requests, each run twice on v0.5.0, followed by 30 targeted regression trials on v0.5.1. All 150 completed. The original early-draft behavior remains visible in the evidence; all 30 regression trials passed the recorded file checks.
Read the results, limits and retained evidence. These are installed Codex trials on gpt-6-astra, not a real-user study or cross-client guarantee.
The complete skill ZIP and SHA-256 checksum are attached. $align remains the command. Follow your client's configured update route in the installation guide.
ALIGNS.md v0.5.0: Ask when context matters
Brief requests now get a context check before clarification. The agent asks about consequential gaps, groups related questions, and proceeds within delegated scope. Repeated misunderstandings can prompt a focused conversation about when the user wants questions.
Task-specific answers remain separate from lasting preferences. Only explicitly requested or confirmed preferences are saved in ALIGNS.md; a method update does not change personal criteria or confirmation dates.
English, Korean, and Japanese documentation is synchronized. The complete skill ZIP and SHA-256 checksum are attached. $align remains the command. See installation and updates for the update route used by your client.
Validation: 31 package tests and Skill Creator validation passed. This release does not claim new benchmark results.
ALIGNS.md v0.4.0
The project is now ALIGNS.md. The repository is https://github.com/hmmhmmhm/aligns.md and the site is https://aligns-md.hmmhmm-hm.chatgpt.site.
- Use
ALIGNS.mdfor project agreements. The$aligncommand is unchanged. - Preserve existing
ALIGN.mdcriteria and confirmation dates when migrating. If both filenames exist, review their differences first. - Update installation and marketplace references using the migration guide.
- Refresh English, Korean, and Japanese documentation, banners, social previews, and benchmark display names. Archived benchmark evidence remains unchanged.
Validation: skill package and setup tests, benchmark harness tests, archived evidence verification, website checks, and desktop/mobile browser inspection. Plot values match the previous release. This release does not claim a new benchmark run.
The ZIP contains the complete portable align skill. Native plugins use the repository manifests. Custom-domain connection for aligns.md is pending.
v0.3.0: History-informed alignment
Review selected history to clarify what good work means.
- Choose recent or full PR feedback, selected dates, and accessible project conversations after seeing the scope and estimated time.
- Collect every available PR feedback page, report partial coverage, and resume unfinished PRs after interruption.
- Parse selected Codex or Claude JSONL text with project boundaries and explicit unsupported-format reporting.
- Use taste invariants as scoped, revisable criteria supported by evidence. Check exceptions and confirm inferred criteria before saving them.
The helpers make no model calls and do not rewrite ALIGN.md. Raw evidence stays in a chosen private working file outside Git. Time estimates disclose their assumptions. Supported transcript formats are documented; this release does not claim universal client compatibility.
Validation: 30 package tests, Skill Creator validation, synthetic collection/parser fixtures, and independent Astra review with regression fixes. A live metadata-only GitHub plan was checked; no personal logs were used.
The ZIP contains the complete align skill. Verify it with SHA256SUMS.txt. Native plugin users can update through their existing installation route.
align v0.2.0: Ongoing personal alignment
ALIGN now supports ongoing personal alignment through a maintained skill.
- Add focused review after important judgment mismatches, recurring missing criteria, uncertain direction, and before long autonomous runs.
- Add native Codex and Claude Code plugin marketplaces, a static Claude session reminder, and a safe instruction-link helper.
- Replace the standalone README prompt with an installation request that connects the project and sets up a supported update policy.
Use native automatic updates where supported, or an agreed targeted refresh when alignment starts. Method updates never rewrite personal criteria or confirmation dates.
Validation: 11 package/helper checks, independent scenario exercises, local installation and a remote targeted upgrade preserving a personal agreement. Native plugin runtime activation and automatic refresh have not been exercised for this release.
The ZIP contains the complete portable skill and MIT license. Native plugins install from the repository manifests.
align v0.1.0
First public release of the align skill.
Create and maintain ALIGN.md through focused conversations about a user's goals, priorities, preferences, and decision criteria. Supports new and existing projects, Git-aware review, one-off exceptions, and distinct full and partial confirmation timestamps.
Install
npx skills add hmmhmmhm/align.md --skill alignThe ZIP contains the self-contained skill folder, MIT license, and optional Codex interface metadata. Extract it into your agent's supported skills directory for manual installation.
Validation
- Skill structure validated.
- Local and GitHub installation verified for Codex, Claude Code, and Cursor using skills CLI 1.7.2.
- Seven behavioral scenarios and one repeated-review follow-up executed by independent acting agents.
- Behavior evidence and limitations are documented in docs/validation.md.
Installation is not a guarantee of automatic execution on every task. Read docs/usage.md for instruction integration and docs/distribution.md for verified listings and pending registrations.
MIT License. Copyright (c) 2026 hmmhmmhm.