Skill Loader

GITHUB_USERNAME: your-github-username REPO_NAME: your-skills-repo SKILLS_REPO_URL: https://raw.githubusercontent.com/your-github-username/your-skills-repo/main/README.md SKILLS_REPO_API: repos/your-github-username/your-skills-repo PERMANENT_SKILLS: skill-loader, skill-cleanup (add any others here, e.g. a brand-reference skill)

Load only the skills you need for this session from the central skill repository.

Step 1 — Fetch the catalog

Run this to get the README (substitute your configured SKILLS_REPO_API):

bash -c 'source ~/.bashrc && gh api {SKILLS_REPO_API}/contents/README.md --jq ".content" | base64 -d'

Read and understand the full catalog of available skills and their descriptions.

Step 2 — Collect project context signals

Check each of the following (skip silently if not present):

# What files exist in the working directory?
ls -la

# Package manager / language signals
cat package.json 2>/dev/null | head -30
cat requirements.txt 2>/dev/null | head -10
cat go.mod 2>/dev/null | head -5
cat Cargo.toml 2>/dev/null | head -5
cat composer.json 2>/dev/null | head -10

# Project instructions
cat CLAUDE.md 2>/dev/null
cat .claude/CLAUDE.md 2>/dev/null

# Infrastructure signals
cat Dockerfile 2>/dev/null | head -20
cat docker-compose.yaml 2>/dev/null | head -20
cat docker-compose.yml 2>/dev/null | head -20

Also note:

• What task or goal has the user stated for this session?
• What file extensions dominate the working directory?

Permanent residents (always available — never fetch, never add to manifest)

These skills (defined in PERMANENT_SKILLS above) are installed locally at all times. Do not recommend fetching them and do not write them to .session-skills:

(Replace this list with your configured PERMANENT_SKILLS values)

• skill-loader
• skill-cleanup

Step 3 — Generate recommendations

Based on the catalog descriptions and project signals, select skills that are directly relevant to this project and session. For each recommendation, give one sentence of reasoning tied to a specific signal. Skip any permanent residents — they are always available without fetching.

Aim for 3–8 skills. Be selective — do not recommend everything. Skills you don't recommend are still available if asked.

Step 4 — Present and confirm

Present in this format:

Based on [key signals observed], I recommend fetching these skills:

✓ `skill-name` — reason tied to specific signal
✓ `skill-name` — reason tied to specific signal
...

Also available but not recommended for this project: skill-a, skill-b, skill-c

Confirm this list, adjust it, or add skill names?

Wait for the user to respond before fetching anything.

Step 5 — Fetch approved skills

For each confirmed skill name, run (substitute your SKILLS_REPO_API):

# Fetch SKILL.md
bash -c 'source ~/.bashrc && \
  SKILL=<skill-name> && \
  mkdir -p ~/.claude/skills/$SKILL && \
  gh api "{SKILLS_REPO_API}/contents/$SKILL/SKILL.md" --jq ".content" | base64 -d > ~/.claude/skills/$SKILL/SKILL.md && \
  echo "Fetched: $SKILL/SKILL.md"'

Then fetch any reference files for that skill:

bash -c 'source ~/.bashrc && \
  SKILL=<skill-name> && \
  REFS=$(gh api "{SKILLS_REPO_API}/git/trees/main?recursive=1" --jq ".tree[] | select(.path | startswith(\"$SKILL/references/\")) | .path") && \
  if [ -n "$REFS" ]; then
    mkdir -p ~/.claude/skills/$SKILL/references
    echo "$REFS" | while IFS= read -r path; do
      fname=$(basename "$path")
      gh api "{SKILLS_REPO_API}/contents/$path" --jq ".content" | base64 -d > ~/.claude/skills/$SKILL/references/$fname
      echo "Fetched: $path"
    done
  fi'

Repeat for each skill in the confirmed list.

Step 6 — Write the session manifest

cat > ~/.claude/skills/.session-skills << 'MANIFEST'
<skill1>
<skill2>
<skill3>
MANIFEST

Write one skill name per line, listing every skill that was fetched in this session.

Step 7 — Confirm

Report which skills were fetched and are now active. Remind the user:

When you're done with this session, invoke skill-cleanup to remove these skills and keep ~/.claude/skills/ lean.