Headshot Skill

Generates professional AI headshots of the principal with Tools/Headshot.ts, which calls the Gemini image model named geminiImage in LIFEOS/TOOLS/models.ts. Matches facial structure from approved reference photos (which ones go in is the CreateHeadshot rule), and defaults to 2K resolution and 1:1 square (3:4 for portrait). A real photoshoot is slow and expensive to redo every time the look needs to change, and generic AI image tools drift off the actual face; anchoring generation to approved reference photos keeps the output recognizably the subject, at the size and ratio each use (LinkedIn, profile pics, publication bios) needs.

Ideal State

• Every headshot is recognizably the subject, because its face comes from approved real reference photos whose energy matches the requested expression.
• Each output lands at the size and ratio its use needs and is saved in the subject's USER data, never in the skill tree.
• Nothing reaches the principal as a first draft, and nothing public ships until the principal has picked it.

Intent Routing

Intent	Trigger	File
CreateHeadshot	"headshot", "profile picture", "LinkedIn photo", "professional photo", "update profile pic"	Intents/CreateHeadshot.md

File	Read when
References/ReferenceLibrary.md	choosing references: the approved pool, scene plates, energy-matching rules, growing the pool
Setup.md	first-time setup, or Headshot.ts check exits non-zero
Tools/	Tools/headshot.help.md holds the full CLI flag reference

Data & Config

Everything about the subject lives in <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Headshot/Config.md (configRoot is $CLAUDE_CONFIG_DIR when set, else ~/.claude): the directory of approved reference photos with a sentiment for each, the default look (attire, backdrop, mood), and optional scene plates. This skill body is generic and ships no photos or looks; first-time setup is Setup.md.

• Tools/Headshot.ts check verifies the config, that the listed photos exist, and that GOOGLE_API_KEY is set. Run it before any generation; when it exits non-zero, stop and report what it printed, which points at Setup.md.
• Generated headshots (the subject's face) are written to LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Headshot/generated-headshots/, never into the skill tree. API keys come from the env file: GOOGLE_API_KEY (from <configRoot>/.env).

Quality loop

Nothing this skill produces reaches the owner as a first draft. The artifact goes through Skill("ArtDirection"): a direction council before new or large work, then the cross-vendor critique panel (skills/ArtDirection/Tools/Critique.ts) scores it against the rubric and names defects, then this skill iterates on those defects until the bar or reports the shortfall. Hand-over is skills/ArtDirection/Tools/Deliver.ts, which refuses without a passing critique. Record the owner's verdict with skills/ArtDirection/Tools/Calibrate.ts.

Gotchas

• Consent. Use only photos of yourself or of people who agreed to have their likeness used. Never generate a headshot of anyone else from their photo without their consent.
• Generation never writes to ~/Downloads/ (ArtDirection, 2026-09-17). Output goes to generated-headshots/; a copy reaches ~/Downloads/ only through Deliver.ts after a passing critique. This replaces the 2026-03-22 rule to output there first, as all Art outputs once did.
• Edits inherit the base image's energy. An expression-only edit keeps the mood of the image it starts from, so a tired-looking base (a paused video frame, say) stays tired through any number of expression prompts. If the base's energy is wrong, no prompt fixes it: add a reference photo that already has the energy you want (for example one tagged with an upbeat sentiment in your config) and let it own the face.
• Edit vs regenerate tradeoff. Full regeneration invents the scene and tends to look generated; strict single-image edits inherit the base's flaws. What usually works is a composite: an approved reference for the face, a real background plate, and an instruction to place the subject in that scene.
• Likeness for video thumbnails or posters must match the footage. Details visible in the video (clothing, accessories, grooming) and a background taken from the video itself should be checked against real frames, not approved by eye alone. If your install has a thumbnail or video workflow that checks likeness against the footage, use it for this case.

Examples

Example 1: LinkedIn Profile Update

User: "Generate a headshot for my LinkedIn profile"
→ CreateHeadshot
→ A critiqued 2K 1:1 headshot in the default look, delivered to ~/Downloads, original kept under generated-headshots/

Example 2: New profile image for a public site

User: "Make me a new profile picture for the site"
→ CreateHeadshot
→ A labeled pick sheet of 2-4 critiqued variants; nothing on the site changes until the principal picks one

Example 3: Specific Style Request

User: "I need a headshot with professional lighting, 3:4 ratio"
→ CreateHeadshot
→ A 3:4 portrait-format headshot with the requested lighting, delivered after a passing critique