For AI coworkers

Use ConvergeQA from Codex, Claude Code, Cowork, OpenClaw, Hermes, or similar AI coworker environments.

ConvergeQA can be used by a person working with an AI tool. Install a local skill first, then let the tool prepare the call, preserve packets and bundles, and keep spend decisions visible.

Compare

Ask 5 models the same question and get their responses plus a synthesis. Best for fast but thorough research.

Iterate

Let ConvergeQA propose changes, then accept or decline each recommendation before finalizing the document.

Critique

Audit a finished deliverable without editing the original. Best for PDFs, presentation files, and final drafts.

Human-driven vs. agent-driven

Human-driven means the AI coworker prepares and submits the call, but a person chooses the mode, approves spend, reviews recommendations, and decides what to accept.

Agent-driven means the AI coworker has an explicit decision policy and may continue rounds without stopping after every packet. Use this only when the stop condition, round limit, and decision policy are written down before the first live call.

The safe setup pattern

1. Keep keys out of chat.
Store your key as an environment variable. Do not paste it into prompts, screenshots, logs, commits, or documents.
2. Install a local skill.
Copy one of the starter SKILL.md templates below into your AI coworker's skills folder. Prompts alone are not enough for multi-step packet, decision, and export workflows.
3. Show the setup first.
Ask the AI coworker to show mode, tier, model panel, input source, synthesis choice, and whether it will pause for decisions.
4. Use approved content only.
Do not submit PHI, patient identifiers, secrets, private credentials, or regulated data.
5. Save the evidence.
Keep the packet, bundle, decision log, model panel, and credit use with the work record.

Keys and names

Developer API key: use this for human-driven AI coworker calls. Generate it on the Developers page, store it as CONVERGEQA_API_KEY, and send it as X-API-Key.

Org service-account key: use this for delegated organization workflows. Org admins create it on the Organization page. Store it as CONVERGEQA_SERVICE_ACCOUNT_KEY, send it as X-Service-Account-Key, and expect it to start with cqa_sa_. Service-account keys are org-owned, revocable, credit-controlled, and scoped to approved Compare, Critique, Iterate, packet-read, bundle-export, and delegated-decision paths.

The environment variable name does not change the key type. The key type is determined by where the key was created. If a key came from the Developers page, it is a Developer API key even if your local variable was named something else.

Do not paste any key into the AI conversation.

Local skill templates

Use these as review-ready local skill drafts for Codex, Claude-style tools, Cowork, OpenClaw, Hermes, compatible IDEs, or MCP clients. Each template tells the AI coworker to use ConvergeQA's supported clients, keep credentials in environment variables, pause at the right decision points, preserve exact ConvergeQA output, and export evidence before cleanup.

Compare skill

---
name: convergeqa-compare-agent
description: Use ConvergeQA Compare for a prompt, question, draft, or document. Do not use for Iterate or Critique.
---

# ConvergeQA Compare Agent

Use this skill when the user asks to compare a prompt, question, draft, document, source packet, model answer, or research direction across multiple models. Do not use this skill for revision loops or finished-deliverable audits.

## Non-negotiables

- Use the supported Compare due-diligence client when available:

tools/agent_clients/compare_due_diligence_cli.py

- Prefer the client, wrapper, or background worker provided by the project over browser automation.
- Keep credentials in environment variables. Developer API keys use CONVERGEQA_API_KEY and X-API-Key. Org service-account keys use CONVERGEQA_SERVICE_ACCOUNT_KEY and X-Service-Account-Key on approved private agent paths.
- Do not submit PHI, secrets, private credentials, regulated data, or unapproved customer data.
- Send the same prompt to the panel by default. Use specialist prompts only when the active runner explicitly supports that mode and the user approves the routing.
- Ask exactly: "Synthesis desired? Y/N?" unless synthesis was already approved. If the answer is no, use the no-synthesis flow.
- Return exact model outputs and exact ConvergeQA synthesis before adding any assistant summary.
- Export or save the packet JSON and bundle ZIP when those artifacts are available.

## Setup to show before running

Show a concise setup card before the first live call:

- Mode: Compare
- Title or short input label
- Input source: pasted text, local file, URL, or prepared packet
- Template or prompt framing
- Tier: Premium unless the user asks for Budget
- Model panel and count
- Prompt routing: same prompt or approved specialist prompts
- Synthesis choice: ask the required Y/N question if unknown
- Credential kind: Developer API key or org service-account key
- Spend or credit boundary
- Artifact output folder
- Sensitivity check: confirm no PHI, secrets, private credentials, regulated data, or unapproved customer data

## Workflow

1. Prepare the input outside the chat when possible, especially for long files.
2. Show the setup card and wait for approval if any spend, synthesis, sensitive-source, or model-panel choice is unclear.
3. Run the supported Compare client.
4. Poll until the packet is complete, failed, or stopped by the configured limit.
5. Save the packet JSON and bundle ZIP when available.
6. Read the saved artifacts, not just the terminal summary.
7. Present the exact per-model responses first.
8. Present the exact ConvergeQA synthesis if generated.
9. Then, if useful, add a clearly labeled assistant summary.
10. Close with artifact paths, cost or credit metadata when reported, and any limits.

## Closeout

Report:

- Compare mode used
- Model tier and panel
- Whether synthesis was requested and generated
- Packet and bundle locations
- Cost or credit metadata when reported
- Any failures, missing model responses, or unverified claims

ConvergeQA records a multi-model due-diligence process. Do not claim it proves the content is true, legal, medical, complete, or production-ready.

Iterate skill

---
name: convergeqa-iterate-agent
description: Use ConvergeQA Iterate to improve or revise a draft through recommendation-by-recommendation decisions. Do not use for Compare or Critique.
---

# ConvergeQA Iterate Agent

Use this skill when the user asks to improve, revise, iterate, strengthen, tighten, or converge a draft through recommendation-by-recommendation decisions. Do not use this skill for side-by-side model comparison or final audit only.

## Non-negotiables

- Use the supported Critique/Iterate client when available:

tools/agent_clients/service_account_reviews_cli.py --mode iterate

- Prefer the client, wrapper, or background worker provided by the project over browser automation.
- Keep credentials in environment variables. Developer API keys use CONVERGEQA_API_KEY and X-API-Key. Org service-account keys use CONVERGEQA_SERVICE_ACCOUNT_KEY and X-Service-Account-Key on approved private agent paths.
- Do not submit PHI, secrets, private credentials, regulated data, or unapproved customer data.
- Default to human-driven decisions. Pause after each recommendation packet.
- Use agent-driven decisions only when the user has approved a written decision policy, stop condition, and round limit before the first live call.
- Render ConvergeQA recommendations exactly before asking for accept or decline decisions.
- Do not rewrite the draft yourself as a substitute for the ConvergeQA decision flow.
- Export or save the final packet and bundle before cleanup.

## Setup to show before running

Show a concise setup card before the first live call:

- Mode: Iterate
- Draft title
- Input source: pasted text, local file, URL, or prepared packet
- Template or review lens
- Tier: Premium unless the user asks for Budget
- Model panel
- Reference material or style constraints
- Desired length, read time, or output shape
- Iteration control: one round, continue until approved, or approved round limit
- Decision control: human-driven or approved agent-driven policy
- Spend or credit boundary
- Artifact output folder
- Sensitivity check: confirm no PHI, secrets, private credentials, regulated data, or unapproved customer data

## Decision grammar

Use simple decisions that can be logged:

- A [number] means accept that recommendation.
- D [number] means decline that recommendation.
- A all means accept all currently shown recommendations.
- D all means decline all currently shown recommendations.
- Notes without an explicit accept are treated as comments or decline instructions, not silent acceptance.
- If the user gives ambiguous instructions, ask for clarification before submitting decisions.

## Workflow

1. Prepare the draft and references outside the chat when possible.
2. Show the setup card and get approval for unclear spend, source, model, or decision-control choices.
3. Start the Iterate run with the supported client.
4. Poll until the recommendation packet is available.
5. Read the packet from saved artifacts when possible.
6. Present each ConvergeQA recommendation exactly, with stable numbering.
7. Collect decisions using the decision grammar.
8. Submit the decision packet.
9. Continue only within the approved iteration control.
10. Finish by exporting or saving the final packet and bundle.

## Closeout

Report:

- Iterate mode used
- Model tier and panel
- Rounds completed
- Decision log summary
- Final packet and bundle locations
- Cost or credit metadata when reported
- Any recommendations skipped, declined, or left unresolved

ConvergeQA records the revision decision process. Do not claim the result is certified, legally approved, medically safe, or complete unless an appropriate human reviewer has separately confirmed it.

Critique skill

---
name: convergeqa-critique-agent
description: Use ConvergeQA Critique to audit a finished deliverable without editing the original. Do not use for Compare or Iterate.
---

# ConvergeQA Critique Agent

Use this skill when the user asks to audit, critique, review, red-team, or assess a finished deliverable without editing the original. Do not use this skill for ordinary revision loops or model comparison.

## Non-negotiables

- Use the supported Critique runner or Critique client when available:

tools/agent_clients/agent_review_critique_runner.py
tools/agent_clients/service_account_reviews_cli.py --mode critique

- Prefer the client, wrapper, or background worker provided by the project over browser automation.
- Keep credentials in environment variables. Developer API keys use CONVERGEQA_API_KEY and X-API-Key. Org service-account keys use CONVERGEQA_SERVICE_ACCOUNT_KEY and X-Service-Account-Key on approved private agent paths.
- Do not submit PHI, secrets, private credentials, regulated data, or unapproved customer data.
- Treat the original deliverable as unchanged unless the user separately asks for an Iterate pass.
- Render the full ConvergeQA audit output exactly before asking for decisions.
- Pause for human decisions unless agent-driven authority was approved with a written policy, stop condition, and round limit before the first live call.
- Export or save the audit packet and bundle before cleanup.

## Setup to show before running

Show a concise setup card before the first live call:

- Mode: Critique
- Deliverable title
- Input source: pasted text, local file, URL, or prepared packet
- Deliverable type: PDF, presentation file, document, page, or other
- Template or audit lens
- Tier: Premium unless the user asks for Budget
- Model panel
- Reference material or requirements to check against
- Decision control: human-driven or approved agent-driven policy
- Spend or credit boundary
- Artifact output folder
- Sensitivity check: confirm no PHI, secrets, private credentials, regulated data, or unapproved customer data

## Decision checkpoint

After showing the full audit, ask what to do next:

- Agree: accept listed findings as-is.
- Disagree [finding numbers]: record disagreement or mark findings as not accepted.
- Go deeper [finding numbers]: request more analysis on specific findings if the runner supports another round.
- Continue: proceed within the approved policy or round limit.
- Finish: export the audit artifacts and stop.

Treat unlisted findings as agreed only when the user explicitly says so. Otherwise keep decisions literal and logged.

## Workflow

1. Prepare the finished deliverable and references outside the chat when possible.
2. Show the setup card and get approval for unclear spend, source, model, or decision-control choices.
3. Start the Critique run with the supported runner or client.
4. Poll until the audit packet is available.
5. Read the packet from saved artifacts when possible.
6. Present the exact ConvergeQA audit, including synthesis and finding structure.
7. Ask for decisions using the decision checkpoint.
8. Continue only within the approved decision control.
9. Finish by exporting or saving the audit packet and bundle.

## Closeout

Report:

- Critique mode used
- Model tier and panel
- Audit packet and bundle locations
- Decisions recorded
- Cost or credit metadata when reported
- Any unresolved findings, skipped exports, or unverified claims

ConvergeQA records an audit process. Do not claim it certifies the deliverable as true, legal, medical, complete, or publication-ready unless an appropriate human reviewer has separately confirmed it.

Quick prompt starters

Compare

Use ConvergeQA Compare on the text I provide. Use Budget tier unless I say Premium. Send the same prompt to each model. Ask me whether I want synthesis before running. Save the packet and bundle paths. Do not put keys in chat or command arguments. Do not submit PHI, secrets, private credentials, or regulated data.

Iterate

Use ConvergeQA Iterate on this draft. Before running, show me the setup: title, template, model tier, panel, reference material, length target, and decision control. Pause after each recommendation packet so I can accept or decline changes. Export the final bundle after I finish. Do not submit PHI, secrets, private credentials, or regulated data.

Critique

Use ConvergeQA Critique on this finished deliverable. Treat the original as unchanged. Before running, show me the setup: title, template, model tier, panel, reference material, and whether decisions are human-guided or delegated. Show the full audit before asking for decisions. Export the audit bundle after finish. Do not submit PHI, secrets, private credentials, or regulated data.

Evidence checklist

Open Web Interface API details for developers