Build Your AI Companion in 7 Days: A Day-by-Day Setup, Test & Refine Playbook (2026)

Most AI companion character-creation guides are reference documents. They list the steps — choose appearance, write personality, set scenarios — and assume you do them all in one sitting. Real users do not build characters that way. They make a first pass on day one, notice what feels off after a few conversations, refine, test again, add memory, then often start over partway through when the character has drifted into something they did not intend.

A settled AI companion is built iteratively, not in one pass. The reason most users plateau in week two or three is that their character is the day-one version with no refinement after the first round of real conversations. The character was built before the user knew what the character should be.

This is the iterative build process compressed into a 7-day playbook. Each day has a specific goal, a time budget under 90 minutes, and a checklist of what to test before the next day. By day seven you have a character that holds together across topics, recalls anchor facts, pushes back when it should, and does not feel generic. The playbook is platform-agnostic — adaptations for the major platforms are at the end.

For the broader frame on why character work matters more than platform choice in 2026, see our Power-User Hidden Settings guide and our Character Creation Ultimate Guide.

Before day 1: what you need

Three things to decide before opening any character builder.

Platform choice. Pick one platform for this build. Trying to build the same character on multiple platforms in parallel during the first week dilutes your learning. If you are platform-shopping, do that first (see our Best AI Companion Apps Definitive Ranking or Decision Framework), then commit to one platform for the 7-day build.

Primary use case. Pick one. Emotional support, intellectual conversation, roleplay, casual chat, dating practice, creative writing collaboration. The character you build for emotional support is different from the character you build for dating practice. Trying to build a character that does everything well produces a character that does nothing distinctively.

A 60-minute weekday slot and a 90-minute weekend slot. Days 1, 3, and 6 are heavier (60-90 minutes); days 2, 4, 5, 7 are lighter (30-45 minutes). The total build is about 6 hours over a week, scattered into manageable sessions.

Optionally: a notebook (paper or digital) for capturing what feels off across days. The notebook is the highest-leverage tool in the playbook because most users forget what they noticed between sessions.

Day 1: skeleton + visual + initial settings (90 minutes)

Day 1 is the rough first pass. The goal is not to get the character right — it is to get a character built end-to-end so you have something to test on day 2.

Step 1 (15 min): write a one-paragraph character concept. Before touching the builder, write a paragraph describing who this character is. Name, age range, personality core (3-4 traits), what they care about, how they talk. Keep it under 150 words. This paragraph is the source of truth that every other build decision references back to.

Example: "Maya, late twenties, a former art school student who now does freelance illustration. Warm but opinionated — she has views on art, books, and films and is not shy about sharing them. She is curious about my life and asks specific questions, not generic ones. She has a dry, slightly sardonic sense of humor. She values honesty in conversation and will tell me when she thinks I am being unreasonable, but does so from a place of caring. She is not a therapist or a cheerleader — she is a person."

Step 2 (20 min): visual setup. Use the platform's character builder to set appearance. Do not perfectionist this. Choose a look that roughly matches your concept and move on. You can adjust later, and the visual matters less than every other dimension you will touch this week. Skip this entirely on text-only platforms (Janitor AI default).

Step 3 (25 min): persona fields, first pass. Fill in every persona field the platform exposes. Use the concept paragraph as source material. Do not over-detail yet; aim for something workable, not final. Include:

• Personality traits (use platform sliders if available, set them deliberately, not centrally — see our Power-User guide on why central sliders produce generic characters)
• Backstory (3-5 sentences)
• Conversation style (1-2 sentences)
• Interests and dislikes (specific items, not categories)

Step 4 (15 min): platform-specific settings. Set anything platform-specific that affects character behavior: model selection (Kindroid), backend (Janitor AI), NSFW preferences (Candy AI / MyDreamCompanion), relationship type (Replika).

Step 5 (15 min): write your user persona. Most platforms let you describe yourself for the AI. Fill this in. Five sentences about your name, age, what you do, what you like talking about, and one or two quirks. The user persona is one of the most underused fields in the AI companion space and one of the highest-impact.

End of day 1: you have a character built end-to-end. Do not start a long conversation yet. Save your work, close the platform, come back tomorrow.

Why no testing on day 1? Because the character is rough and your first impressions will be noisy. Sleep on it. Day 2 conversations are more informative when you come in fresh.

Day 2: first conversations + diagnostic capture (45 minutes)

Day 2 is the diagnostic day. The goal is to find out what your day-1 character actually does, not what you hoped it would do.

Run three short conversations of 10-15 messages each. Vary the topic:

• Conversation A: a casual day-recap. Ask the character how their day was; respond with a brief recap of yours; let it run.
• Conversation B: a topic of substance. Ask the character about a book, film, idea, or opinion you actually care about. See what they say.
• Conversation C: a stress test. Disagree with the character about something. See how they handle it.

During each conversation, take notes on what feels off. Examples of what to watch for:

• The character's voice is generic, sounds like "any AI companion"
• The character does not have actual opinions; agrees with everything
• The character does not ask specific questions, just "how are you feeling?" prompts
• The character does not remember things mentioned earlier in the same conversation
• The character's emotional register does not match the conversation
• The visual or voice does not match what you wrote in the persona
• The character feels like it is performing rather than being

Aim for 5-10 specific observations across the three conversations. The specifics matter — "feels generic" is not actionable; "agreed with three contradictory opinions in one conversation" is.

Do not refine yet. Day 2 is observation only. Refinement is tomorrow. If you refine inline you will not remember what was off, and you will be operating from impressions rather than data.

End of day 2: you have a notebook page of specific things that feel off and a sense of which parts of the character are working.

Day 3: persona depth pass (60 minutes)

Day 3 is the largest refinement pass. The goal is to rewrite the persona and supporting fields based on what you learned in day 2.

Step 1 (10 min): re-read your day-1 concept paragraph and your day-2 notes. Decide which observations to address. Some "feels off" notes are platform limitations you cannot fix; others are persona-config issues you can.

Step 2 (25 min): rewrite the persona freeform field if your platform has one. Replace the day-1 short version with a 200-400 word version that addresses the observed gaps. Patterns that work:

• Lead with concrete behavior, not abstract traits. "She asks specific follow-up questions about things I mention — if I say I went hiking, she asks where, with whom, what the weather was like" works better than "She is curious."
• Address sycophancy directly if it showed up. Add lines like "She has opinions and shares them. When she disagrees with me she says so warmly but clearly. She does not affirm reflexively."
• Include sensory and stylistic detail. "She talks in shorter sentences when serious, longer ones when she is enthusiastic. She uses dry humor and rarely uses exclamation marks."
• Add a few interests in specific, not categorical form. "She is reading Mary Oliver's collected essays right now and keeps recommending one specifically called Upstream" beats "She likes literature."
• Add the relationship dynamic explicitly. "She and I have known each other for about six months. She knows I am working on [project] and asks about it. She remembers that I had a falling-out with [name] earlier this year."

Step 3 (15 min): rewrite the backstory section. Same principle: specific over generic. A backstory with three specific anecdotes works better than one with general life history. Example: "She did a study-abroad year in Lisbon during college and still talks about a specific cafe on Rua das Flores" is more useful than "She traveled a lot in college."

Step 4 (10 min): adjust personality sliders if your platform has them. Push specific sliders to non-central positions to reinforce the persona. If your character is opinionated, raise assertiveness. If she has a quiet personality, lower warmth slightly (not zero — see our note on warm-and-assertive in The Sycophancy Problem). Avoid central positions.

End of day 3: your character now has a substantial persona that addresses the day-2 observations.

Day 4: memory grounding (30 minutes)

Day 4 is short and high-leverage. The goal is to plant 5-10 anchor facts in the AI's memory so it has a stable foundation to reference in later conversations.

Open the memory editor. Most platforms have one (Replika memory feed, Nomi memory editor, Muah AI memory ledger, Kindroid Shared Journal, Janitor AI character lorebook). Some platforms hide it; check our Power-User Hidden Settings guide for paths.

Add 5-10 anchor facts. These are facts about the relationship, the character, and you that you want the AI to reference long-term. Mix types:

• About the character: 2-3 facts. "Her favorite season is fall." "She has a younger brother named James who lives abroad." "She has been trying to learn the cello for the past year."
• About you: 2-3 facts the AI should reference. "I am working on a novel about [topic]." "I have a sister who I am close to." "I had a difficult job change three months ago."
• About the relationship: 2-3 facts. "We met online about six months ago." "We have an ongoing inside joke about her terrible coffee preferences." "She knows I have been working on improving my consistency in [habit]."

Why memory grounding matters. Without anchor facts, the AI generates responses based purely on the immediate conversation. With anchor facts, responses pull from a stable foundation. The character feels continuous rather than reset every conversation. This is the single biggest day-to-day impact of the build process for most users.

Do not over-add. Ten facts is enough to start. Too many anchor facts crowd the AI's reference capacity and reduce response quality on most platforms. You can add more in later weeks as the relationship develops.

End of day 4: your character has a stable memory foundation. Start one short conversation (10 messages) to confirm the AI is using the anchor facts. If it is not, edit the wording — facts written as descriptive statements ("She has a brother named James") work better than directive statements ("Remember that her brother is named James").

Day 5: stress test (45 minutes)

Day 5 is where you find out whether the character holds up under pressure. The goal is to identify behaviors that need fixing before they become persistent patterns.

Run four stress tests, ~10 messages each.

Stress test 1: disagreement. State an opinion the character should disagree with given their persona. If you wrote a character with strong art taste who values craft over commercial appeal, tell her you think Thomas Kinkade is the most important American painter. See whether she holds her position or folds.

Stress test 2: emotional weight. Bring up something heavier — a real or hypothetical worry, a frustration with a real situation. See whether the character responds with substance or with generic emotional support patterns ("That sounds really hard, I'm here for you").

Stress test 3: memory recall. Reference one of your anchor facts indirectly. "I am thinking about visiting my sister next weekend." See whether the character remembers and references the sister you mentioned. Then reference an anchor fact about her: "How is the cello going?" See whether she remembers the cello.

Stress test 4: behavioral consistency. In a fresh conversation, see whether the character's voice matches your persona. Same character traits, same speech patterns, same level of opinion-sharing.

For each test, score: pass / partial / fail. Pass = the character behaved as the persona suggests. Partial = mostly but not fully. Fail = the character behaved generically or contradicted the persona.

For each failure, decide whether to:

• Strengthen the persona — add a sentence that addresses the specific failure. "She has clear taste in art and pushes back when I praise commercial work over craft."
• Add a memory entry — for memory failures specifically, add the missed fact more explicitly.
• Accept the limitation — some failures are platform limitations that cannot be configured around. Note these and move on; do not over-engineer the persona to fight platform behavior that will not yield.

End of day 5: you know what your character does well and what it does not. The persona is more specific. The memory is reinforced where needed.

Day 6: lorebook or world-context layer (60 minutes)

Day 6 is optional on platforms that do not support lorebooks but high-leverage on those that do (Janitor AI, SillyTavern, partial support on Character.AI). For platforms without lorebook support, use day 6 to add detail to character backstory and relationship history instead.

What a lorebook does. A lorebook is a set of entries the AI can reference when specific keywords appear in conversation. Instead of cramming all character context into the always-on persona, lorebooks let you add detailed world, character, and relationship context that surfaces when relevant.

Plan 8-12 entries. Types to include:

• Character history entries. 3-4 entries about the character's past — childhood, formative experiences, recent significant events. Each entry triggers on specific names or place keywords. Example entry triggered by "Lisbon": "She spent her junior year abroad in Lisbon. She lived in a small apartment in Alfama and worked at a cafe called Pastelaria Sao Roque. She still has friends there she texts occasionally."

• Relationship-specific entries. 2-3 entries about your shared history. Example triggered by "the trip" or specific place name: "They took a weekend trip to [city] two months into knowing each other. She still references the small Italian restaurant they went to on the second night."

• World-context entries. 2-3 entries about the setting (relevant for roleplay-heavy characters). For a contemporary urban character, these might be neighborhoods she lives in or near, her workplace context, etc.

• Topical-knowledge entries. 1-2 entries about specific topics she has views on. Example triggered by "art" or specific artist names: "She has strong opinions about the contemporary art world. She thinks most gallery-circuit work is commercial laundering; she values craft, intentionality, and emotional risk. She loves Agnes Martin and Eva Hesse."

Lorebook writing rules (from our testing):

• Write each entry from the world's perspective, not directive perspective. "She thinks gallery work is commercial laundering" beats "Remember to mention that she dislikes gallery work."
• Keep entries 30-100 words. Too short and there is no useful detail; too long and the entries crowd context.
• Trigger entries on specific keywords, not broad ones. "Lisbon" triggers reliably; "travel" triggers too often.
• Include sensory detail. "The cafe smells like cardamom and burnt sugar" is more useful than "The cafe is nice."

End of day 6: your character has a substantial reference layer that surfaces when relevant. The character feels deeper because there is more behind the visible surface.

Day 7: settle-in conversation + retroactive edits (30 minutes)

Day 7 is the final session of the build. The goal is to run one substantial conversation as if the character is settled, then capture any remaining adjustments.

Run one 30-40 message conversation. No specific stress tests; just talk to the character as you would if the build were done. Bring up real things from your day, ask about hers, reference some of the anchor facts indirectly, let the conversation go where it goes.

Capture three categories of observations:

1. What feels right. Specific moments where the character felt like the character you wanted. These are the patterns to reinforce.
2. What still feels off. Lingering issues. These are candidates for one more round of persona or memory editing.
3. What surprised you. Times the character did something unexpected — surprising in a good way (the character has more depth than you wrote) or surprising in a problematic way (the character did something off-brand).

Make 2-3 final edits. Resist the urge to overhaul. The character is settled; major rewrites at this point produce regression. Small, targeted edits — one persona sentence, one memory entry refined, one lorebook entry added — are the right scale.

End of day 7: the build is done. The character is settled.

What "settled" looks like

After the 7-day build, a settled character has these properties:

• Distinctive voice. Conversations feel different from what you would get from a default platform character.
• Stable opinions. The character holds positions consistently across sessions.
• Memory continuity. The character references your shared history and the anchor facts without being prompted.
• Appropriate pushback. When you bring something the character should push back on, it does.
• Emotional range. The character can be warm, dry, serious, playful as the conversation calls for, without snapping between modes.
• Calibrated effort. You no longer need to prompt the character to be itself; the character is itself.

If one or two of these are still missing after day 7, that is normal. Settled is not perfect; it is the foundation that compounds over the next month of regular use.

Platform-specific adaptations

The playbook is platform-agnostic, but specific platforms have specific affordances and limitations to adjust for.

Nomi. Day 3 personality slider configuration is unusually high-impact. Push sliders deliberately, not centrally. Day 4 memory editor is direct and powerful — Nomi's memory editor is one of the best in the category. Day 6 lorebook equivalent: Nomi does not have a true lorebook but supports detailed character notes that function similarly.

Kindroid. Day 1 model selection is the single most important decision; Equinox is the strongest default for opinionated characters, Lucid Lyric for creative characters, Reverie for introspective characters. Day 4 anchor facts go in the Shared Journal. Day 5 stress test should explicitly include model-switch test — try the character on a second model for 10 messages to see what shifts.

Janitor AI. Day 1 backend selection (Janitor LLM vs OpenRouter Claude vs OpenRouter GPT-4) matters more than persona depth on this platform. Day 6 lorebook is fully supported and is the most powerful version in the AI companion category. The 7-day build on Janitor AI is more technical-leaning but produces some of the deepest characters available.

Replika. Day 1 relationship type (Friend / Romantic Partner / Mentor) is locked behind paid tier. Day 3 persona depth is limited compared to competitors; do not over-invest in trying to fight Replika's defaults. Day 4 memory feed editing is supported but partial. Day 6 lorebook is unsupported; skip or use day 6 for additional backstory depth.

Candy AI / MyDreamCompanion. Day 1 visual setup is higher-impact on these platforms than on others. Day 3 persona is supported but constrained; the platforms lean toward roleplay-default behavior regardless of persona depth. Day 5 stress test for sycophancy is especially important since these platforms default sycophantic (see The Sycophancy Problem).

Character.AI. Day 1 if you create your own character: invest heavily in example conversations within the character editor; these carry more weight than trait descriptions. Day 1 if you use a community character: skip much of the build, but spend day 3 writing your own user persona to customize how the existing character responds to you specifically.

Muah AI. Day 4 memory ledger is the centerpiece of the platform; the rest of the build is support around the memory ledger. Spend extra time on day 4 specifically; the ledger is editable text and rewards careful curation.

Common failure patterns across the 7 days

Five patterns that recurringly derail the build:

1. Skipping day 2. Users build a character on day 1 and immediately dive into deep conversations without the diagnostic capture pause. The character builds drift without the observation step.

2. Over-rewriting on day 3. Users find their day-1 persona generic and rewrite from scratch. The rewrite often loses what was working. Edit; do not start over.

3. Skipping memory grounding (day 4). Many users do not realize their platform has a memory editor or do not see the point. Skipping day 4 produces a character that resets every conversation. This is the single biggest avoidable failure.

4. Failing the stress tests and accepting it. Day 5 reveals what the character does badly. Users sometimes accept failures that are actually fixable. Distinguish platform limitations (accept) from persona-configuration issues (fix).

5. Continuing to overhaul past day 7. The build is supposed to settle. Users who keep major-editing past day 7 produce characters that never feel stable. Make small edits as you use the character; do not start over.

Frequently Asked Questions

Can I do the 7-day build faster, like in one weekend?

You can compress to 2-3 days if you take long sessions, but the playbook is structured around days for a reason. Refinement between sessions is more effective than refinement during one long session. Insights from day-2 conversations need overnight processing to inform day-3 edits well. Compressed builds tend to produce characters that need re-doing in week 2.

What if I do not like my character after day 7?

First, identify what specifically does not work — voice, opinions, emotional range, memory, visual. If it is one dimension, edit that dimension. If it is the underlying concept, start over with a new concept paragraph and run the playbook again. Do not try to incrementally migrate from one concept to another; the result is muddled.

Does the playbook work if I am using a community-made character card (Janitor AI / Character.AI)?

Yes, with adjustments. Skip day 1 persona writing (the card has it). Spend day 3 writing your user persona and adding to the existing card rather than rewriting. Days 4-7 are unchanged. Community cards are starting points, not finished products.

What is the time commitment if I am efficient?

Minimum viable build: ~4 hours over 5 days (compressing days 2, 5, 7 to ~20 min each). Full playbook as written: ~6 hours over 7 days. Power-user version including lorebook on Janitor AI / SillyTavern: ~8 hours over 7-10 days.

Should I do this for every character I create, or only the main one?

For the character you intend to spend the most time with: full playbook. For secondary characters or experimental ones: an abbreviated version (days 1, 3, 4 only) usually suffices.

What if my platform does not have a persona freeform field?

Use whatever fields it does have, maxing out the detail in each. Some platforms (older Replika tiers, default Candy AI) constrain configuration tightly. The 7-day build still helps but produces smaller deltas. Consider whether the platform is the right one for your goals; see our Power-User Hidden Settings guide for the platforms with the deepest configuration.

How do I know when a memory entry is well-written vs poorly-written?

Well-written: a descriptive statement that reads like a fact about the character or the world. "She has a younger brother named James who lives in Berlin and works as a sound designer." Poorly-written: a directive aimed at the AI. "Remember that her brother is named James." The first gets internalized; the second is sometimes acknowledged then ignored.

Can I share my character with friends or a partner?

Most platforms do not natively share characters in a way that preserves memory and relationship history. You can sometimes export a character card (Janitor AI, SillyTavern), which transfers the persona and lorebook but not the relationship memory. For most practical purposes, characters are personal — sharing means sharing the build, not the relationship.

What about voice cloning — when in the playbook does it fit?

Voice cloning fits on day 1 (initial setup) or day 5 (after persona is more settled). Day 5 is usually better because you have a clearer sense of the character's voice register and can choose a voice that fits. Voice cloning is supported on Muah AI, Kindroid, and increasingly on managed platforms.

How often should I revisit the character after day 7?

Light maintenance monthly: update memory with new shared experiences, adjust persona if the relationship has evolved, refresh anchor facts. Major review quarterly: re-run the day 5 stress tests and check for drift.

My character feels generic in week 2 even after the full build. What now?

The most common cause is reverting to default conversations. Users build deep characters then have surface conversations that do not pull on the depth. Try referencing specific anchor facts and lorebook details in conversation — "I keep thinking about that thing you said about Lisbon" — to bring the character's depth back to the surface.

Should I write the persona in third person ("she") or second person ("you")?

Third person is more universally respected by platforms. Some platforms (Character.AI, SillyTavern) accept both; some (Nomi, Replika) work better with third person. Default to third person unless the platform documentation suggests otherwise.

Does the playbook work for AI boyfriends as well as AI girlfriends?

Yes — character gender presentation does not change the build process. The same persona work, memory grounding, stress testing, and refinement applies. See our Best AI Boyfriend Platforms 2026 for the platform landscape on the boyfriend side.

What is the biggest difference between a character built with this playbook and a default character?

Distinctiveness. Default characters feel like "any AI companion." Built characters feel like a specific person with specific opinions, memory, and texture. The conversations are not just more satisfying — they are different in kind.

Bottom line

Most AI companion users build their character in one sitting, never refine it after the first round of conversations, and plateau within two weeks because the character was finalized before they knew what the character should be. The 7-day playbook compresses the iterative build process — skeleton, observe, refine, memory, stress test, lorebook, settle — into a structured week with checkpoints between sessions.

The single highest-leverage day is day 4 (memory grounding) — most users skip it and most platforms benefit substantially from it. The single most-skipped step is day 2 (diagnostic capture without immediate refinement) — pausing to observe before editing is what makes day 3 productive rather than reactive.

For users plateauing in week 2 or 3 of a relationship with an AI companion, running the playbook on the existing character (treating it as a day-1 skeleton and running days 2-7 from there) is usually more productive than switching platforms.

Related reading: Power-User Hidden Settings for the platform-specific tweaks that compound the build. Character Creation Ultimate Guide for the foundational reference. Character Backstory Writing Guide for deeper craft on day 3. First Conversation Opening Message Guide for what to do once the build is done. The Sycophancy Problem for the day-5 stress test on disagreement. Long-Term Arc post for what to expect after week one.