Guizang's agent skill turns your repo's recent updates into a finished promo video — real components, code-rendered animation, an original code-written soundtrack — no video editor involved.
- agent-skill
- video
- promo
- marketing
SureForge
An instruction-only skill that makes agents research before they ask, plan before they build, and verify before they call anything done.
SureForge encodes a working procedure for complex tasks as plain text: four phases — research and clarify, plan, execute, deliver — each ending in a gate that returns READY, REPAIR or BLOCKED. Three tiers scale the ceremony: Light for small reversible fixes, Standard with four gates and two complementary checks each, and Full mode requiring three verification methods from the agent plus three chosen freely by a fresh-context independent reviewer at every gate, with at most three review rounds before a gate is declared BLOCKED instead of shipped. There is no runtime, hook or dependency — the agent follows it like any other skill, and small tasks are meant to stay small (a typo fix triggers no research project). The package ships a real evaluation kit: thirteen failure scenarios with pass/fail oracles, twenty activation prompts with expected tiers, and five synthetic benchmark tasks. The author's pilots report GLM-5.2 through Devin following the workflow in 12 of 13 scenarios across 38 sessions, and Grok 4.6 at maximum effort across 24 runs declaring its tier and reporting missing reviewers instead of pretending checks happened.
What it does
- Four phases with hard gates (READY / REPAIR / BLOCKED): research before asking, plan before building, verify before delivering, independent review before done
- Three tiers — Light, Standard, Full — so a typo fix stays a typo fix while high-stakes work gets three agent-chosen plus three reviewer-chosen verification methods per gate
- Skipped questions are not answers: every finding must be confirmed, refuted with evidence, or marked unresolved before anything changes
- Ships an evals kit — 13 failure scenarios, 20 activation prompts, 5 benchmark tasks with hidden grading — plus package tests and a mutation audit
When to use it
High-stakes or hard-to-reverse agent work — migrations, deletions, releases, security fixes — where "the agent said it was done" is not good enough, and teams that want a written, tiered procedure instead of vibes.
Getting started
# with the Skills CLI (Node.js 22.20+), from your project: npx skills add Da7-Tech/SureForge npx skills add Da7-Tech/SureForge --agent claude-code -y # skip the prompt # manual: copy the whole skills/sureforge/ folder (SKILL.md + references/ + assets/) # Claude Code: .claude/skills/sureforge/ or ~/.claude/skills/sureforge/ # verify the package from the repo root (Python 3.11+): python3 -B -m scripts.check_package python3 -B -m unittest discover -s tests -v
Things to know
Instruction-only means no enforcement: a weak model can still skip gates, which is exactly what the pilots measured rather than solved. The published pilots used small synthetic tasks chosen to test adherence, not to show gains on real underspecified work — the README says a three-arm comparison on real task sets is the next step. The Skills CLI may write a skills-lock.json with local paths; inspect it before committing. Platform paths were verified at v1.0.0 for Claude Code, Codex, Cursor, Devin and Hermes — hosts change their directories, so re-check the current docs if a skill is not discovered. MIT.
Repository
Da7-Tech/SureForgeCharted
Related on the map
geo-sleuth
SkillAn agent skill that geolocates a photo with nothing readable in it — OSM geometry, elevation skylines and satellite tiles turn “where was this taken?” into a measured answer with an error radius.
- agent-skill
- geolocation
- osint
- openstreetmap
Birdview
SkillA skill that makes AI coding agents map the architecture and declare the blast radius before they touch a line of code — rendered as one standalone HTML page.
- skill
- architecture
- code-visualization
- change-scope