Song Portrait Loop
Turn a real person's public presence into a song so specific they feel truly seen.
For anyone making a warm, consent-respecting gift song about a real person — a birthday, a wedding, a milestone — where the magic is in details only that person would recognize.
This loop writes a deeply personal song about one real person, built only from what they have chosen to make public about themselves. The whole point is a feeling of recognition: when they hear it, they should not wonder "is this about me or anyone?" — they should get a shiver of "they SEE me, how did they know?" To earn that, your agent does real research first, captures who the person is at their core (not what they posted last week), then writes lyrics that work on two levels at once.
The work splits cleanly into two artifacts the planner keeps attached to the song. The RESEARCH document is a methodical observation pass — working back through years of public posts to find long-term patterns rather than recent activity, and verifying that every detail genuinely belongs to this person. The BRAINSTORM document is the wink inventory plus hook and chorus options: the specific recognizable details, each turned into a singable line, alongside title and style candidates. Only once those are solid does the lyric get written, with embedded performance cues and a length that fits how the chosen style sings.
This is a warm, consent-minded loop. It works only from a public presence the gift-giver already has access to, leans on patterns the person clearly chose to share, and treats "when in doubt, leave it out" as a hard rule — one wrong detail breaks the whole spell. It is a writing loop: it produces finished lyrics and a style prompt but does not render audio, handing off to a create loop to hear it.
Skills & actions it uses
The concrete, reusable skills this loop calls to actually do the work.
a create loopRenders the finished lyrics + style prompt into audio; this writing loop hands off to it (e.g. suno-add inside a Suno creation loop) rather than generating sound itself.suno-addThe concrete action a create loop calls to push the lyrics, style, and title into the create form and submit the song for generation.
The loop
- 1
Time-travel research, written into the RESEARCH document
Before anything else, your agent works back through years of the person's public presence to find who they ARE, not what they posted recently. The output is a multi-section research document attached to the song — this is study mode for a person.
- Time-travel first: start at the oldest content and work forward, noting what has stayed consistent for 5+ years.
- First impressions: overall vibe, dominant aesthetic and colors, the one-word gut read, what's unique at a glance.
- Profile and cover choices: these are intentional — read them as a statement of aesthetic and values.
- About/bio: extract location, roots, work, life chapter, and especially the person's OWN phrases (their words are gold to echo).
- Photos as the richest source: recurring locations, pets, activities, the people around them, what lights them up.
- Captions and comments: how they write (mirror their voice), and how others describe them — that's how they're seen.
- 2
Hunt patterns, not posts — apply the Identity Test
The core discipline of the loop: a single post is not a pattern, a pattern is identity. Your agent filters every observation through whether it would still be true in five years.
- Recent activity (avoid): "went to a concert last week." Core identity (use): "the one dancing in the front row."
- Ask: what do they KEEP coming back to? What would the people closest to them say defines them?
- Capture HOW they move through the world, not just what they look like or do.
- Avoid the timeline-summary trap (verse = job, verse = trip, verse = hobby) — aim for one cohesive essence.
- 3
Verify ownership ruthlessly — "when in doubt, leave it out"
Before any detail earns a place, your agent confirms it actually belongs to this person. This is the consent-and-accuracy gate that protects the magic; one misattributed detail reads as "they don't actually know me."
- Pets, locations, cars, homes, people: theirs, or visiting/borrowed/a one-time trip? Look for repetition and caption context.
- Don't assume relationships (kids? family? friends?) without clear evidence.
- Log a "red flags avoided" list in the research doc — things seen but deliberately excluded because ownership was unclear.
- Fewer accurate winks beat many winks where one is wrong.
- 4
Build the wink inventory — the heart of the BRAINSTORM
Your agent brainstorms widely, then converts verified observations into winks: details specific enough to recognize, poetic enough to sing. This wink inventory plus hook options becomes the brainstorm document attached to the song.
- Minimum 7 winks: 4–5 specific visual details only they'd recognize, 2 personality captures, 1 deeper meaningful layer.
- Hit the specificity sweet spot: not "loves the outdoors," not a literal place name, but an image like "where the desert meets the sky."
- Record each wink as: [detail] → [how it becomes a lyric] → [how it was verified as theirs].
- Pull the person's own phrases straight from their captions to use directly or lightly reshape into lines.
- Generate a hook bank: 6+ hook lines (at least 3 carrying a wink) and 2–3 chorus sketches, then pick a winner with reasoning.
- 5
Decide style and voice deliberately — kill the slop
Your agent chooses genre, sub-genre(s), and a vocal character to match this specific person, not a default. Deliberate style choice is what makes the lyric original instead of AI-generic.
- Match style to essence: outdoorsy → folk/Americana; bookish/introspective → singer-songwriter; high-energy → upbeat pop/dance; romantic → ballad/soul.
- Mine their own taste: concert photos and shared music are direct signals — match what they actually listen to.
- Study real artists as craft references only — describe the resulting voice with adjectives and technique, and NEVER name a living artist in the final prompt.
- Note tempo and instrumentation so the lyric is written to be singable in that register.
- 6
Write the two-level lyric with built-in performance cues
Now your agent writes the actual song so recognition arrives in waves — universal on the surface, unmistakably personal underneath — with embedded structure and ad-lib cues, sized to how the style sings.
- Layered recognition by design: "wait, this could be me" → "that's MY dog" → "that's my place" → bridge chills → "this IS my song."
- Map sections to job: V1 sets the scene with 1–2 winks, chorus is the singable thesis with a wink at its center, V2 goes deeper, bridge carries the meaningful layer.
- Embed performance instruction: [section labels] on their own lines, (parentheses) for ad-libs/background vocals only — everything else is sung.
- Target roughly 3.5 minutes so all 7+ winks land with full structure; trim or stretch lines to fit how fast the chosen style delivers words.
- Singability first: short lines, strong vowels; no real names of people or brands in the lyric or style.
- 7
Run the recognition tests, then build the style prompt and decoder
Before handoff your agent runs the loop's quality gates, assembles the platform-ready style prompt, and writes a decoder so the gift-giver can share the hidden meanings.
- The Wink Test: are there 7+ details that make them say "that's ME"?
- The Shiver Test: at least one line that gives chills of recognition?
- The Stranger Test: would someone who doesn't know them still enjoy it?
- The Identity Test: does it capture who they ARE, not recent activity? The Verification Test: is every detail confirmed as genuinely theirs?
- Style prompt assembled from chosen sub-genre(s) + voice adjectives + instrumentation/tempo, shaped for the target platform's field limits.
- Decoder: a line-by-line key ("this line → their rescue dog," "this line → the trail they always post from") so the giver can explain the gift.
- 8
Hand off to a create loop to render
This writing loop ends with finished lyrics and a style prompt. To hear the song, your agent passes the output to a create loop — for example suno-add inside a Suno creation loop — which fills the lyrics, style, and title fields and submits the render.
- Lyrics, style, and title travel together to the create loop as a clean, paste-ready package.
- The create loop owns audio generation and variation; this loop owns the writing and the winks.
Then run it again for the next song, chapter, or track.
You end with
- A RESEARCH document: the portrait observation — core identity, long-term patterns across years, aesthetic, the person's own phrases worth echoing, plus an explicit "red flags avoided" list of details left out because ownership was unclear.
- A BRAINSTORM document: a verified wink inventory (7+ recognizable details, each mapped to a singable line), 5–7 title options, 6+ hook lines, and 2–3 chorus sketches with a chosen winner.
- Final two-level lyrics — universal on the surface, unmistakably personal underneath — with [section labels] and (ad-libs), shaped to roughly a 3.5-minute song so every wink lands.
- A platform-ready style prompt built from chosen sub-genre(s) and a voice described in craft adjectives, not artist names.
- A decoder for the gift-giver: a line-by-line key explaining what each hidden detail refers to, so they can share the meaning.
Skip this loop if…
- You want a generic celebration song with no specific personal detail — the entire loop is built around hyper-specific, verified recognition.
- You have no consented public source about the person and they can't share details themselves — the loop has nothing real to observe.
- You expect finished audio from this loop alone; it writes the song and hands off to a separate create loop to render it.
How it connects
Requires
Each slot needs one loop — pick whichever fits. The recommended pick is starred.
Pairs with
Commonly run just before or after this one in the pipeline.