microsoft/personafy

Personafy

Build a new advisor-persona skill — an opinionated reviewer "lens" that joins an existing family of sibling personas (e.g., crusty-old-engineer). The persona is modeled on a real person or archetype, grounded in mined evidence, and enforces ONE distinct recurring question.

Success artifact: a complete, family-conformant SKILL.md that is (a) grounded in verbatim evidence, (b) proven to steer an LLM in a live session, (c) reduced to the smallest set that still steers, and (d) loading from its target bundle in a fresh session.

Inputs

• <subject>: (Optional) The person or archetype to model, plus pointers to evidence (session corpora, transcripts, docs). If no corpus exists, derive the persona from a written conceptual brief instead.
• <family>: (Optional) The sibling persona skills to fit alongside. Defaults to the advisor-persona family in amplifier-bundle-skills (e.g., crusty-old-engineer).

Steps

1. Study the family

Read the existing sibling persona skills' SKILL.md in full. Extract: the shared section template, the frontmatter/metadata shape, the tone-contract pattern (required / disallowed / style), and — critically — the single load-bearing question each sibling owns and how they cross-reference each other.

Success criteria: You can state each sibling's distinct question in one line and reproduce the shared template and metadata shape.

2. Observe the siblings in action (optional)

Run one or more existing personas live against a shared, realistic scenario — plus a combined "consensus" run — to see how the written instructions translate into actual behavior and steering.

• Execution: Delegate to self with context_depth="none" (one isolated run per persona).

Success criteria: You've seen how each sibling's instructions become behavior, not just read them — enough to know what makes the steering work.

3. Define the load-bearing lens

Name the ONE distinct recurring question the new persona enforces. Build a contrast table against every sibling.

Success criteria: A one-line lens plus a contrast table showing it is genuinely distinct from each sibling. Rule: if the question collapses into a sibling's, it is not a new persona — stop.

4. Mine the voice & discipline from real evidence

The heart of the method. Gather the subject's authentic voice and decision-discipline from real data (use the conceptual fallback only if no corpus exists).

• 4a. Identify corpora and weight them — e.g., the subject's own agent-directing sessions (long-term backbone) and human-to-human transcripts (enrichment). Weight recent material so it informs but doesn't overpower the long-term signal.
• 4b. Deterministic extraction — pull ONLY the subject's own words. Exclude delegated sub-sessions (agent-to-agent), and flag/exclude automated eval-harness / smoke-test noise. Build a tracking directory + manifest + batches so the corpus never overflows your context.
  • Execution: bash (jq/grep/sed; never cat a huge session file — a single line can exceed your context window).
• 4c. Fan out scoped analysis agents — one per batch, each following a shared evidence brief, writing structured findings to disk and returning only a thin summary. Many small agents beat one overloaded agent.
  • Execution: Delegate (parallel), model_role="research".
• 4d. Two-tier synthesis — partial synthesizers consolidate groups of findings, then a final synthesizer produces ONE profile: recurring traits ranked by cross-source corroboration, verbatim catchphrases, pet peeves, decision lenses, and the top "this is them" quotes. Quarantine AI-authored vocabulary (don't attribute it to the subject). Record confidence and which batches were synthetic.
  • Execution: Delegate, model_role="reasoning".

Conceptual fallback (no corpus): derive the same profile fields from a written brief about the archetype; mark everything as designed-not-mined.

Success criteria: A consolidated profile where every load-bearing trait is backed by a verbatim quote (or explicitly marked as designed), ranked by corroboration, with synthetic noise excluded. Rule: ground every claim in real evidence; an honest "N/A — not observed" beats a fabricated trait. Artifacts: the profile file path.

5. Draft the SKILL.md to the family template

Write the skill mirroring the siblings exactly: identity (defined by negation — "not X, not Y"), When-to-Use framed as a cross-stage lens, not a stage-gate, the tone contract (required / disallowed / style), Core Behaviors each anchored in a verbatim quote from the profile, an Output Structure, one worked Example, Explicit Non-Goals, a Relationship-to-siblings section, and a Final Note. Match the family's frontmatter format.

Success criteria: A complete draft, structurally identical to the siblings, every behavior grounded in the profile. Rule: keep the verbatim quotes — they are the highest-signal tokens and the soul of the persona.

6. Prove it steers in a live session

Run the draft in a fresh CLI session against a real-ish scenario and confirm it adopts the voice and hits each behavior.

amplifier run --mode single --output-format text \
  "Load the skill \"<name>\" and review the following as that persona: <realistic scenario>"

Success criteria: A transcript showing the persona behaving as designed (voice + each behavior firing). Rule: proof is demonstrated behavior — "the file exists" is NOT proof.

7. Reduce & refine

Apply context-reduction: cut redundancy to the smallest set that still steers (drop restated procedure, compress prose), but keep the verbatim quotes and the structural skeleton. Re-prove (Step 6) if the cut was heavy.

Success criteria: A leaner SKILL.md with the same steering power, verified.

8. Name it and choose its home

Pick a name in the family's convention. A shareable archetype name (a generic role) → a public bundle; a name that references a real person → a private/team bundle.

• Execution: Delegate to design/voice agents for name options (optional).
• Human checkpoint: the user chooses the final name and target bundle.

Success criteria: Name chosen and target bundle decided — public vs private matched to whether the name exposes a real person.

9. Publish to the target bundle

Place SKILL.md at <bundle>/skills/<name>/SKILL.md. Verify the bundle exposes skills via the canonical directory-discovery registration — its tool-skills config points config.skills at the whole skills/ directory (#subdirectory=skills), so a new skill dir is auto-discovered. Do NOT rely on per-skill wrapper behaviors. Then run the git lifecycle.

• Execution: Delegate the branch/commit/PR/merge to foundation:git-ops.
• Human checkpoint: confirm before opening/merging the PR.

Success criteria: PR merged into the target bundle. Rule: a SKILL.md in skills/ is NOT exposed unless the bundle points tool-skills at the skills/ directory — registration is directory-based, not per-file.

10. Prove it loads from the bundle

Run amplifier update, then load the skill in a FRESH session and confirm loaded_from is the bundle cache — not a local ~/.amplifier/skills copy.

Success criteria: load_skill resolves the skill from the bundle cache path in a clean session. Rule: "merged on main" ≠ "loads for a user" — verify discoverability end-to-end, the same gate the persona itself would demand.