38 SKILL.md files · Claude Code, Codex and Gemini · free, no email
Not templates. These are the SKILL.md files our own agents load every day, exported from the live tree. Take the whole set and drop it into Claude Code, Codex, or a Gemini project file.
.zip, 174 KB. No email, no sign-up.
Separated index
Download the SKILL.md, put it where your agent looks for skills, then let the trigger line do the work.
Before you install one, read our Claude Code best practices for making "done" mean proven: the checks, the Stop hook and the numbers behind these skills. For the hook itself, field by field, see Claude Code hooks and the one that refuses "done".
Downloadable Markdown
Each card links to a Markdown file carrying the trigger, the procedure, the quality gate, and the failure that made the rule necessary.
Write a new Claude Code skill the way Anthropic's own guidance says to write one: sharp trigger, tight scope, real examples.
Scaffold a Claude Code subagent with its own system prompt, tool allowlist, and handoff contract.
Build Claude Code hooks that fire on tool events, so a rule is enforced by the harness instead of remembered by the model.
Save, list, and reconnect to previous Claude Code sessions instead of losing a long-running thread.
Call Claude Code and Codex from Node with primary/fallback routing, schema-validated JSON, and graceful degradation.
Force the laziest solution that actually works: does it need to exist, is it already here, does the stdlib do it, can it be one line.
A four-file planning system (Prompt, Plan, Implement, Documentation) that gives a project durable memory across sessions.
Stop shooting once. Change, run, verify the real result, compare to the previous round, keep only clear wins, stop at a verified plateau.
Turn a rough idea into a written plan plus an ordered, dependency-aware task file a runner can execute.
Execute a multi-part task file autonomously, part by part, with self-healing, testing, and progress reporting.
LLM-as-judge evaluation built for determinism: binary rubrics, golden sets, and alignment checks on the judge itself.
Integration-test an endpoint whose happy path costs minutes and dollars, without ever firing the expensive pipeline.
Watch a log, detect fixable errors, deduplicate by signature, and dispatch an autonomous fix run for each new one.
Find and fix duplicate side effects caused by timeout-plus-retry against a non-idempotent API.
Atomic read-modify-write for processes running in cluster mode, where two workers race the same JSON or row.
Trace unwanted punctuation in non-Latin scripts across every layer that touches the string.
Diagnose the Claude CLI 'Killed' / SIGKILL crash caused by bloated session transcripts, and fix it without losing history.
Reroute a feature endpoint that returns a partial artifact back through the platform's real orchestration path.
Enforce hard business constraints across an LLM planner, a media generator, and a branded composer without trusting any stage to remember them.
Anti-AI-slop design: build, audit, or redesign a page so it does not read as a template with the defaults left on.
Apple's interface and motion principles translated to the web: springs, interruptible transitions, materials, optical typography.
Reverse-lookup glossary that turns 'the bouncy thing when a popover opens' into the exact term a designer or model will recognise.
The quality gate for visual work: open the rendered artifact, check crop, hierarchy, spacing, safe zones, RTL behaviour, then revise and look again.
Capture video from anything that renders in a browser and prove it is smooth before it ships, by frame index rather than screen recording.
Generate and edit raster images with reference control over likeness, palette, and composition.
Generate and upscale images and video through a curated registry of best-in-class models instead of one default.
Automated visual QC for generated graphics: capture the frame, have a vision model hunt overlaps, cutoffs, font failures, and missing logos, then retry.
Multi-source research with citation tracking, persisted evidence, and a structured report you can audit later.
Optimise content to be cited by AI answer engines: AEO, GEO, llms.txt, knowledge bundles, agent-readable structure.
Turn one asset and one honest number into an overnight optimisation loop: one change per round, score it, keep winners, revert losers.
Extract clean markdown from a web page, dropping navigation and clutter, so a read costs a fraction of the tokens.
Strip the tells of AI writing: inflated symbolism, hedged conclusions, rule-of-three padding, the em-dash habit.
Communication frameworks for outbound messages: specificity, honest stake, mirrored language, one low-burden ask.
Markdown to a genuinely well-set PDF, including Hebrew and RTL that keeps code, paths, and tables the right way round.
Create and edit JSON Canvas files with nodes, edges, and groups, so an agent can draw a real diagram.
Obsidian-flavoured markdown done properly: wikilinks, embeds, callouts, properties, tags.
A daily, evidence-first round on a page taking paid traffic: read the analytics, watch real sessions, run one honest test, ship the change.
A hard guard at the final second before any email send, blocking a duplicate that another path or lane already sent.
How to use the files
These files are useful as downloadable assets, but the production value comes from connecting them to your data, permissions, tools, logs, approvals, and evaluation loop.
Download the Markdown file that matches the workflow.
Adapt the inputs, guardrails, data access, and approval rules.
Connect it to Claude, Codex, private tools, APIs, or internal systems.
Run it with human review, then improve from real usage.
Build with Bles
Bles Software builds custom AI agents, Claude and Codex skill systems, internal tools, API integrations, and production automations for companies that need operational leverage.
About Us
Features
Testimonials
Contact Us
© 2026 Bles Software, Yehud-Monoson, Israel. All Rights Reserved.
