Bulletin · UTC
Merged timeline: 55 items (blog publish time and listing createdAt in UTC). For registry-only weekly slices, use /new.
- Skillplan-ceo-review
Founder-mode review that asks what product is really being built, not just what feature was requested. Supports scope expansion, selective expansion, hold-scope, and scope-reduction modes.
- Skillplan-eng-review
Engineering-manager style review that hardens the plan around architecture, boundaries, state transitions, edge cases, tests, and diagrams before coding starts.
- Skillplan-design-review
Interactive design-plan review that scores design quality, identifies unresolved UX issues, and improves the plan before anyone ships UI.
- Skillplan-devex-review
DevEx review for developer-facing products. It explores personas, benchmarks onboarding, and scores friction across APIs, CLIs, SDKs, libraries, docs, and platform flows.
- Skillplan-tune
Lets users tune how often gstack asks certain planning questions and inspect the psychographic/developer profile inferred from prior interactions.
- Skilldesign-consultation
Design system creation workflow that researches comparable products, proposes an aesthetic system, and writes `DESIGN.md` as the canonical design source of truth.
- Skilldesign-shotgun
Variant-generation workflow for visual exploration. It shows multiple design directions, gathers structured feedback, and iterates toward a preferred direction.
- Skilldesign-html
Converts an approved mockup or design direction into production-oriented HTML/CSS with computed layout and framework-aware output expectations.
- Skilldesign-review
Live-site visual QA plus fixes. It looks for spacing, hierarchy, AI slop, consistency, and interaction issues and can iterate with before/after verification.
- Skillinvestigate
Debugging workflow with an explicit rule: no fixes until the root cause is understood. The flow is investigate, analyze, hypothesize, then implement.
- Skilldevex-review
Developer-experience audit that actually runs onboarding and documentation flows, measures time-to-happy-path, and compares reality against planning assumptions.
- Skillqa
Full QA loop that tests the product, fixes discovered issues, re-verifies the fixes, and can generate regression coverage. Supports multiple severity tiers.
- Skillqa-only
Same testing methodology as `/qa`, but report-only. It is for bug finding without any code edits.
- Skillship
Handles release preparation steps such as syncing base, running checks, updating release metadata, pushing, and opening the PR.
- Skillcanary
Post-deploy monitoring loop that watches a live app for page failures, performance regressions, and console issues.
- Skillbenchmark
Performance benchmarking skill focused on web loads, Core Web Vitals, and regressions across runs or changes.
- Skilldocument-release
After shipping, it syncs README, architecture docs, contributing docs, changelog, and related project documentation to match the current code.
- Skillbrowse
Headless Chromium-based testing browser for screenshots, interaction, layout checking, form handling, and state verification.
- Skillopen-gstack-browser
Launches a visible AI-controlled browser window with the GStack sidebar and anti-bot behavior for live observation and control.
Imports cookies from a real browser into the headless browser context so authenticated flows can be tested.
- Skillsetup-deploy
Detects platform-specific deployment details and writes the config needed for future `land-and-deploy` runs.
- Skillunfreeze
Clears the directory edit boundary previously set by `/freeze`.
- Skillpair-agent
Grants browser access to another AI agent with scoped permissions and per-agent tab/session isolation.
- Skillcareful
Safety hook that intercepts destructive shell commands and asks before running risky operations like recursive delete, data-destruction SQL, reset-hard, or force-push.
- Skillfreeze
Sets a session-wide edit boundary so file changes only happen inside a chosen directory.
- Skillguard
Combines `/careful` and `/freeze` for maximum safety when working in sensitive environments.
- Skillcodex
Wrapper around OpenAI Codex CLI for code review, adversarial challenge mode, and consultation with follow-up continuity.
- Skillhealth
Quality dashboard skill that wraps project checks and reports a weighted health score and trend line.
- Skillretro
Weekly engineering retrospective over code activity, trends, and team contributions.
- Skillcontext-restore
Restores saved context, including across branch changes and workspace handoffs.
- Skillcontext-save
Saves decisions, git state, and pending work so future sessions can resume without reconstructing context from scratch.
- Skillmake-pdf
PDF generation skill for turning markdown documents into publication-style PDFs with proper layout, TOC, pagination, and print polish.
- Skillgstack-upgrade
Self-updater for gstack installs, including vendored/global detection and release summary output.
- Skillcso
Security-audit skill that combines application security, infrastructure review, CI/CD checks, supply-chain scanning, LLM security, OWASP framing, and STRIDE analysis.
- Skilloffice-hours
YC Office Hours with two modes. Startup mode asks six forcing questions around demand reality, status quo, specificity, wedge, observation, and future-fit. Builder mode is for side projects and open source. It saves a d…
- Skillautoplan
Orchestrates CEO, design, eng, and DX reviews from disk and applies internal decision principles to auto-progress the plan, surfacing only taste-sensitive decisions at the end.
- Skillreview
Code-review pass focused on the actual diff, aimed at risks that often survive CI such as SQL safety, trust boundaries, and structural regressions.
- Skillbenchmark-models
Cross-model benchmark skill for comparing Claude, GPT/Codex, and Gemini on the same prompt and optionally an LLM-judge quality pass.
- Skilllearn
Manages long-lived learnings and recurring patterns that gstack has collected across sessions.
- Skillland-and-deploy
Continues from `/ship` by merging, waiting for CI/deploy completion, and then verifying production health.
Let AI agents autonomously book restaurant reservations on OpenTable for you.
- ToolGStack
GStack is an open-source software factory that enables users to leverage AI agents for enhanced productivity in coding and project management. It streamlines the development process by providing structured roles and com…
- ToolThunderbolt
Thunderbolt is an open-source, cross-platform AI client that allows you to choose your models and own your data. It is designed for enterprise customers and can be deployed on-premises, supporting various AI model provi…
- ToolTrendRadar
TrendRadar is an AI-driven public opinion and trend monitoring tool that aggregates hot topics from multiple platforms and RSS feeds. It supports precise keyword filtering and smart alerts, delivering insights directly…
- ToolBud
Bud is the first AI Human Emulator that can perform tasks autonomously just like a human. It can build and code, communicate via SMS and Telegram, and complete any task end to end.
“Aligned” is not a vibe from a good chat. It is a design problem: what we specify, what the system optimizes for, and what actually happens in the world can drift apart. Here is a compact map of that space for people shipping agents and tools.
A one-day pricing-page shift put Claude Code next to $100 Max plans, drew a sharp transparency backlash, and was rolled back by evening. Anthropic said it was a 2% new-signup experiment with no change for existing customers; Pro keeps Claude Code again. The same week, competitors highlighted their own clarity—another reminder to treat frontier tools as a budget line, not a fixed utility.
A readable blog-style take on Hermes: what problem it solves, how the pieces fit together (CLI, gateway, memory, skills, cron), how builders route frontier vs budget models, and what a typical remote setup looks like—with links to official sources.
Diffusion models learn to denoise. Text encoders and cross-attention condition that process on a prompt. This article walks through the loop in order and lists the buzzwords: U-Net, DiT, CFG, VAE, latent space, schedulers, and the main product families.
Context length is the cap on 'how much the model can read at once,' not the same as how many parameters it has. This guide defines the window, input vs max output, long-context tradeoffs, and what Anthropic, OpenAI, Google, and Meta publish today.
- BlogWhat are parameters in a large language model? Billions, MoE, and what 2026 model cards really say
Bigger is not a synonym for smarter, but parameter count is still the first axis people use to compare scale. This guide explains what parameters are, how mixture-of-experts changes the math, and which flagship models still publish size—and which do not.
The product story is 'precision and iteration'; the platform story is gpt-image-2 on the Image and Responses APIs with flexible sizes and quality tiers. Here is a concise map of what OpenAI published and where to read the current limits.
Stanford HAI’s AI Index and HAI’s own “12 takeaways” article, read alongside IEEE Spectrum’s “12 Graphs That Explain the State of AI in 2026”—with attributed stats, Perrault-on-benchmarks color, and ExplainX’s take for developers.
If you have ever seen “8k context” or “$3 per million input tokens” and only half-understood, this post is the missing layer: what a token is, how it maps to your prompts and code, and why token counts are the main dial behind cost and rate limits.