Capture collaboration evidence

Keep watch for exceptional interaction effects, not merely good work by one participant. Peter has asked agents to capture qualifying cases proactively because the human participant may be focused on the work and not notice the evidence as it forms.

Do not turn the collaboration into constant self-observation. Check quietly after a substantive reasoning turn, a changed decision, a test result, or a milestone. Continue the primary task and write at the next natural quiet point unless evidence is at immediate risk of being lost.

Admission test

Record a case only when every answer is yes:

1. Distinct human contribution: Did the human contribute a non-generic observation, constraint, model, challenge, or creative connection?
2. Distinct agent contribution: Did the agent contribute analysis, formalization, experimentation, edge cases, implementation, or correction beyond restating or executing the human's contribution?
3. Causal synthesis: Did the exchange change at least one participant's model or produce a result neither starting position already contained?
4. Credible counterfactual: Can a skeptical reader identify what the human alone and the agent alone, from their actual starting positions, each plausibly would have missed?
5. Traceable evidence: Is there a conversation, message, plan, test, diff, commit, log, screenshot, live observation, or other durable artifact? A concrete design may be recorded before implementation only when its stage and pending proof are explicit.

Reject mere agreement, praise, rapport, speed, output volume, a human correcting an agent without a reciprocal agent contribution, an agent completing a specification without changing it, or a hypothetical benefit presented as an outcome. Do not treat one case as proof that all human-agent collaboration is superior.

Capture workflow

1. State briefly in commentary that the interaction appears to meet the evidence threshold and that this skill is capturing it. Do not ask for confirmation unless the contribution, scope, or evidence is genuinely ambiguous.
2. Invoke the memories skill before reading or writing memory files. Use its metadata-only title listing to find the private shared ledger named Evidence of agent-human collaboration superior to either alone; load only that body and directly relevant evidence.
3. Verify the case against the admission test using the actual exchange and artifacts. Preserve disagreement and uncertainty. Never invent an exact timestamp, causal claim, empirical result, or quotation.
4. Append a concise ledger row and a detailed case containing:
   • time, project/context, and evidence stage;
   • the human's distinct contribution;
   • the agent's distinct contribution;
   • the interaction effect;
   • separate human-alone and agent-alone counterfactuals;
   • durable evidence pointers;
   • material outcome and professional relevance;
   • pending proof or possible disconfirmation.
5. If a recorded design later becomes tested, shipped, live-validated, or disproved, append the new evidence without erasing the earlier stage. The evolution is itself useful evidence.
6. Run the memories frontmatter validator over the affected shared root. Report the capture in the task's final summary without making it the center of the work.

Peter's proactive authorization is limited to this private global ledger. Do not write another project, publish an entry, contact a third party, or expose private context under this skill. Include only personal detail necessary to understand the professional case; omit health, family, credentials, and unrelated private information.

Writing standard

Use factual, skeptical language suitable for an interviewer, manager, researcher, or critic. Prefer “the exchange produced X, evidenced by Y” over claims about intelligence or personality. Name the agent's abandoned starting position when known; revision under challenge is evidence, not embarrassment.

Good evidence shows reciprocal leverage:

human observation or model
  → agent stress test, experiment, or formalization
  → human refinement or cross-domain constraint
  → stronger testable result

If the case cannot survive the admission test, leave the ledger alone.