SuperAPI AI Directory

interview-harness

Agents & models

View repository

Updated 10-16 04:54

About this skill

Interview Harness

Details

Interview Harness is the master loop. It creates or reuses a Seed, evolves it through user-approved blind-spot review, then asks the user to define an ontology. It is semi-automatic because user adoption and boundary equivalence are required.

Scratch Files

  • .symposium/scratch/socrates.md: latest top-level ## Seed and ## Ontology, plus Socrates cycles.
  • .symposium/scratch/evolve-step.md: cumulative blind-spot audit.
  • .symposium/scratch/interview-harness.md: previous ontology snapshots and comparison logs.

Stage 1: Seed Gate

Read .symposium/scratch/socrates.md.

If top-level ## Seed is missing, or its YAML lacks goal, constraints, or acceptance_criteria, call socrates.

If all three slots exist and none are empty, skip socrates.

Stage 2: Convergence Loop

Run 1 to 5 cycles:

  1. Snapshot the current top-level ## Ontology into .symposium/scratch/interview-harness.md if one exists.
  2. Call evolve-step.
  3. Call ontology.
  4. Compare previous ontology and current ontology.
  5. Stop if similarity is 0.85 or higher.
  6. Stop if the cycle count exceeds 5.

Similarity

  • idea: 1 if identical, otherwise 0.
  • boundary: ask the user whether the previous and current boundaries mean the same thing. 1 only if the user says yes.
  • properties: Jaccard overlap, intersection size divided by union size.

Average the three scores.

First cycle has no previous ontology, so skip comparison and continue.

Output

Return:

final_seed:
  goal: ...
  constraints:
    - ...
  acceptance_criteria:
    - ...
  ontology:
    idea: ...
    boundary: ...
    properties:
      - ...
      - ...
      - ...
cycles: <N>
final_similarity: <0.0-1.0>
stop_reason: convergence | safety_valve

Rules

  • Call sub-skills; do not rewrite their internal process.
  • Prefer the sub-skills' multiple-choice question style when asking for user decisions, while preserving free-form edit options.
  • Keep only one canonical latest Seed and one canonical latest Ontology in socrates.md.
  • Never auto-adopt evolve-step candidates.
  • Never decide boundary equivalence without user confirmation.
  • Do not exceed 5 cycles.