Skills I didn’t write but use every week

Blocks any implementation until the idea has been talked into an actual design and you have approved it.

The single highest-value one on this list. It stops the failure where you describe a feature in one sentence and get four hundred lines of confident, wrong code.

Turns a spec into a written implementation plan of bite-sized tasks, written for someone with zero context.

The plan is the artefact I can actually review. I can read a plan and know whether it is wrong; I cannot always read a diff and know that.

Works through a written plan in a fresh session with review checkpoints between phases.

Splitting the planning session from the building session keeps the model from quietly rewriting the plan to suit what it already did.

Farms independent tasks in a plan out to subagents inside the same session.

Useful once a plan has genuinely parallel branches. Not worth the overhead on small work.

Runs two or more unrelated tasks at once when nothing is shared between them.

Mostly for research sweeps — several independent questions answered in the time of one.

Writes the failing test before the implementation, every time.

I resisted this one longest and was wrong. A test written after the fact tests what the code does, not what you wanted.

Forces a hypothesis and a reproduction before any fix is proposed.

Without it you get plausible fixes to bugs that were never diagnosed, and a file that slowly fills with dead defensive code.

Requires the actual command output before anything can be called done, fixed, or passing.

“It should work now” is the most expensive sentence in AI-assisted building. This skill deletes it.

Asks for a review against the original requirements before merging.

Catches the drift between what was asked for and what was built, which is a different problem from whether the code is correct.

Makes the model verify review feedback rather than agreeing with all of it.

Left alone, models implement every suggestion including the wrong ones. Agreeableness is not a code-quality strategy.

Presents structured options — merge, PR, or clean up — once work is complete.

Stops half-finished branches accumulating, which is the state most of my repos used to live in.

Sets up an isolated worktree so feature work cannot disturb the current workspace.

Cheap insurance. I would rather throw away a directory than untangle a working tree.

The entry point — teaches Claude to look for a relevant skill before answering anything.

Without it, the rest of the collection sits on disk unused. Skills only help if they get invoked.

Creates, reads and edits Word documents, including tracked changes and comments.

Half of what a marketing team produces still has to leave the building as a .docx. This turns a draft into the file the legal review actually wants.

Builds and edits slide decks, layouts, templates and speaker notes.

Good for getting the structure and the words of a deck right. The design pass is still mine.

Opens, cleans, computes and charts spreadsheets.

Cleaning a messy export by hand is an hour I will never get back. This is the single biggest time saver on the list.

Extracts text and tables, merges, splits, fills forms and runs OCR on scans.

Most competitor research arrives as a PDF nobody can search. This fixes that.

Creates new skills, improves existing ones and runs evals to measure whether the description actually triggers.

The eval half is the part people skip. A skill that never triggers is a file, not a tool.

Finds and enriches people and companies — titles, contact details, funding, tech stack, hiring signals.

Prospect and ICP research without the tab-by-tab grind. The enrichment is only as good as the underlying data, so I verify anything I am going to put in a campaign.