collaboration-pattern helps an AI collaborator pause at decisions that matter, check the facts carrying the decision, and choose a sensible way to continue. It is useful when the goal is unclear, evidence is incomplete, an assumption is expensive, or a new finding challenges the current plan.
It is an installable Agent Skill with a client-neutral core and a small Codex/OpenAI adapter.
Use it when:
- an unverified schema, file set, API contract, or source set could change the work;
- customer workflow or purchasing evidence is missing;
- new evidence contradicts an approved direction;
- a specialist workflow may not be installed;
- a claim such as
all,none, orcompletedepends on an unenumerated set; - a decision is costly or difficult to reverse.
For a small isolated choice that is cheap to reverse, the pattern chooses and labels a conservative default instead of starting an unnecessary investigation.
Before
The plan assumes every record has a string
status, so the importer and tests are designed around that type.
After
The schema controls the data model and validation, but no real record has been inspected. Enumerate one real instance first. Keep
statusasUNKNOWNuntil observed, then update the affected requirement and continue with design exploration.
The example is synthetic. It does not reproduce a private project or dataset.
Clone the repository, then copy only skill/collaboration-pattern/ into the directory where your client discovers skills. Keep the installed directory name exactly collaboration-pattern.
For Codex on Windows PowerShell:
git clone https://github.com/zhqi3/collaboration-pattern.git
Copy-Item -Recurse -LiteralPath '.\collaboration-pattern\skill\collaboration-pattern' -Destination "$env:USERPROFILE\.codex\skills\collaboration-pattern"The package uses the Agent Skills directory format. It was initially tested with Codex. Other clients may require a different install location, adapter, or invocation syntax and remain untested for v0.1.0.
Ask the client to use $collaboration-pattern, for example:
Use $collaboration-pattern before committing this plan. Identify the current decision, which unknowns could change it, what must be checked, and which capability should handle the next step.
Understand the current decision
-> identify unknowns that could change it
-> decide whether a factual probe is required
-> choose an appropriate specialist capability
-> act under that capability's rules
-> verify material claims
-> reconsider when evidence changes
The core routes by capabilities such as discovery, planning, diagnosis, research, review, and verification. Optional client mappings live in specialist-routing.md. If a specialist is unavailable, the Skill uses a small fallback or stops; it does not pretend the workflow exists.
| Evidence | What it covers | Limit |
|---|---|---|
| Behavioral scenarios and rubric | Six decision-boundary behaviors | Scenario design does not guarantee behavior elsewhere |
| v0.1.0 results | Three baseline runs per scenario; three historical candidate runs per scenario at cfc64a8; and a corrected S06 recheck at fde0745 |
Run-specific Codex evidence, not a statistical guarantee; the corrected S06 result is not a current whole-package score |
| Deterministic validator and tests | File map, metadata, references, stale branding, privacy patterns, UTF-8, links, and evidence IDs | Pattern checks cannot prove the absence of every privacy or security issue |
| Test matrix | Exact local checks and observed outcomes | Other Agent Skills clients remain untested |
There is no model call in CI. Behavioral responses are preserved as evidence rather than presented as deterministic tests.
v0.1.0 is a local instruction package plus public evidence and validation. It has no GUI, hosted service, telemetry, automatic model API calls, or marketplace submission. It does not reproduce complete specialist methods inside the core.
The installable unit is skill/collaboration-pattern/. Evaluations, validation, release evidence, and repository documentation stay outside runtime Skill context.
Read CONTRIBUTING.md before proposing behavior changes. Report privacy or security-sensitive problems through SECURITY.md.
MIT licensed. See LICENSE.