Personalise

You drop something in — a URL, a paste, a file path, a tool name, a workflow, a prompt template. Output: a focused recommendation on how to adapt it to your world. Not a generic summary. Not a blind install.

Process

1. Classify the input in one line — skill, tool, article, repo, prompt, workflow, framework, content piece, etc.
2. Pull only the context you need — don't read everything:

- Always skim: the user's goals, mission, and "about me" files (typically the project's main instruction file and any user-profile file) - For voice/content inputs: any tone/style/voice file plus relevant content feedback memories - For tooling/code inputs: tech stack notes, available tools/services file - For placement decisions: workspace structure rules, skill placement rules - Always check the memory index for relevant feedback/project entries - For overlap detection: grep existing skills and tasks/backlog to spot duplicates

1. Assess fit — what's gold for the user, what's mismatched, what's missing, what overlaps with what they already have.
2. Output the recommendation.

Output format

Three sections, tight:

What this is — one line.

Take rate — high / medium / low / skip + one-line reason filtered through the user's stated goals.

Personalisation plan — concrete bullets only:

• Where it lives (which workspace / which agent — apply skill placement rules from the project's instructions)
• What to rename (strip upstream branding, match the user's English variant, swap in the user's canonical terms)
• What to drop (anything that conflicts with voice rules, tech stack, or existing feedback memories)
• What to add (links to existing skills/tasks/agents it should reference)
• Prereqs (env vars, packages, tools, infra)
• Approval gates triggered (e.g. paid generation cost approval; messaging-platform replies; external-system actions)

End with one branching question only if there's a real choice. Otherwise stop.

Lenses to apply

• Workspace fit — shared or one specific agent / sub-project? Never duplicate across folders.
• Naming — strip personal/upstream names, match the user's English variant (e.g. AU vs US spelling), use their canonical terms.
• Voice — apply any voice/tone rules in the project: no AI-smell, no hype, terse, options-first when content involves choices.
• Tech stack — match the user's defaults (module system, framework versions, package manager, test runner, linter). Flag mismatches.
• Goals filter — does adopting this move the user toward their stated goals? If no, flag low or skip.
• Existing overlap — grep first. If overlap, recommend updating the existing skill/task instead of creating new.
• Cost gates — flag any paid-model usage (image / video / LLM APIs) — needs cost approval before run.
• External-system gates — flag if it touches messaging platforms, publishing, git push, or APIs that affect shared state.
• Memory hygiene — only propose memory entries if the input is a preference or feedback rule. Tools, ideas, tutorials don't go in memory.

What NOT to do

• Don't install/apply without the user's go. Recommend first, execute on confirmation.
• Don't expand scope — one input, one personalisation. No pre-prepping siblings "just in case".
• Don't write planning documents unless asked. Chat output only.
• Don't generate the personalised artifact (rewritten skill, adapted prompt, edited script) until the user picks the take rate and confirms placement.