Each skill is a short SKILL.md — the only part that ever enters the model's context — plus a script that gets executed, never read. The clever logic costs zero tokens no matter how many times it's reused.
The first three shrink reads, edits and repeated command output. The rest attack everything else: repeated output, long strings, reassurance reads, mechanical edits, logs, payloads, traces, diffs, search hops, re-derived facts, and unbounded floods.
hashpatch
Edit files by pointing at hashed line anchors instead of quoting whole files.
rerun
Cache and reuse output from tests, builds, and lints across repeat runs.
probe
Inspect a module's API by running it, not by reading its source.
seen
Fold already-shown output blocks into one-line pointers.
alias
Shorten long repeated strings (hashes, UUIDs, paths) into small tokens.
believe
Verify a claim about a file against reality without re-reading it.
refactor
Describe a mechanical multi-file edit in a sentence instead of a diff.
mine
Cluster huge logs into a handful of templated lines with counts.
shape
Print the schema of structured data instead of dumping the payload.
trace
Strip library noise from stack traces down to the user-code frames.
sdiff
Summarize a diff by symbol instead of printing every changed line.
q
Answer structural code questions from a cached symbol index.
blast
Show a symbol's callers and callees before you change its signature.
recall
Persist answers to "where/how is X handled" across sessions.
budget
Track token spend per command and block unbounded floods.
Requires only Python 3. Installs into your global Claude directory and wires up the supporting hooks automatically.
git clone https://github.com/Dextrata/nitro-skills.git
cd nitro-skills
python install.py
See the full README for manual install, flags, and a CLAUDE.md example that makes usage legible to the agent.