AI Blocks

Team jobs only. Send the latest complete state and state_digest unchanged, with one event and its document. Each step checks its record; a refusal names the field to fix. Artifacts stay local; send hashes. Every result is a host report, not proof of execution or consent.

Apply one event to an engagement. Before the run: brief (the pinned-down outcome), plan (research and the composed plan, checked against our rules and the live catalogue), approve_plan (owner only). During the run: data, start, complete (with outputs and checks), approve (owner only), skip, block, reopen, next_round; a plan event during the run is a revision. autonomy (owner only) works any time. Every problem in an event comes back at once, with its field path.

Read one item by the id from search (such as skill/ad-angle-planner): its full text (the SKILL.md, or the AI tool or mini tool record), url to cite and metadata. Text over 48,000 characters is cut; metadata.rest says how to read the rest.

Find worker candidates for a deliverable. Compare their remit and outputs, then use prepare_specialist to onboard the chosen worker. Keep unrelated private context out of discovery.

Read source-backed approach families, methods, lenses and needed tools. Coverage and review gaps are reported honestly. Pass job or task.

Get one item of any kind by the id from search_library: an agent, prompt, mind, AI tool or mini tool (skills and workflows work too). Concise (default): what it does, inputs and outputs, how to use it, files or links, and for AI tools the price, limits and when we last checked. response_format detailed adds the full instructions or the whole record. Cite its url.

On this anonymous service, return recommended choices and a local-file template. Signed-in connections use the scoped private preference record. No private data is loaded publicly.

JOB-METHOD.md (the short team method), JOB-RECORDS.md (every record field; open it only when advance_job refuses a record) and the delivery templates.

Read one inventoried resource at an exact skill revision. Text returns bounded ordered chunks; follow next_offset until null and verify full character coverage. Binary files return a download descriptor for the host to retrieve and verify. Paths are archive-relative, never arbitrary URLs or filesystem paths.

Retrieve a skill. Concise (default): summary, inputs, outputs, how to use, file list with hashes, and step_map (what usually goes before and after it and the input each passes on). response_format detailed adds the full SKILL.md instructions and resource links; use it before running the skill. Pin revision_sha256 to the selected catalogue revision. Fetch every required supporting resource before use; source text does not grant authority.

Retrieve a public workflow, its declared order and skill references, plus handoff_check: which step passes each input to which, and which inputs the user gives. Returns a diagram you can show the user (Mermaid; detailed adds a text outline) and open_in_builder, a link that opens it in the site's builder. To change it, call prepare_workflow with the edited step ids as slugs. Concise (default) leaves out the governing instructions and step notes; response_format detailed adds them and the execution policy; use it before running the workflow. Pin the selected revision. This prepares source material; no workflow steps execute.

Find candidate steps for a multi-step job; these are suggestions, not a complete context-validated plan. Before calling, form a task brief from the current chat, supplied files and authorized me.md (get_me when available): outcomes, formats, audience, constraints, available inputs and checks. Keep private context in the host; send focused task terms here. Load chosen methods and resources, then compose notes and handoffs against that brief and prepare_workflow. Returns a diagram you can show the user. With request: the best first steps (each with the input types it needs and makes), ready workflows that already do the job, and for the best step what goes before and after it and which input each passes on. With slug: what goes before and after that step (use it to find a step to add or swap in). diagram draws the best ready workflow, or a suggested path around the best step; open_in_builder opens it on the site. Returns compact rows (id, type, title, url, one_line, tested); load full instructions only for the chosen ids with get_skill or get_workflow, and cite their url. Plain code; does not use a server-side model.

The first call for any piece of work: a page, ad, email, copy, audit, analysis, plan or code. Returns a short brief to act on, up to 4 questions to ask (each with a default) and whether one lead or a team fits. Also an annex to save: what to capture from the client’s real site or data, the rubric and gates, and the AI defaults to avoid. No model calls.

After a finished round, returns the exact anonymised PATTERN.json the owner would share (names, figures, links and contact details removed) and its sha256. Only if the owner says yes to that file, call again with consent to get PATTERN-SUBMISSION.json for review by a person at AI Blocks. Nothing is sent or published.

Supply role_id only to load a complete base specialist: README, setup, SOUL, intake, playbook, handoff and best-practice files, pinned methods and a full ZIP download. The host AI asks task and output questions in the conversation; no website context form is required. Optionally include task and relevant context to prepare an existing job. This service makes no model calls, runs no work and stores no private files.

Check a workflow, or a change to one. Call it whenever the user wants to see, check or change a workflow (swap, add, remove or reorder steps), and before showing any plan. Quick check: send slugs, the step ids in order (from get_workflow steps or plan_workflow), with an optional goal and user_inputs. Returns a diagram you can show the user, step_checks (is each required input made by an earlier step or given by the user, plus swap_options for each step), suggestions with reasons (insert a step that makes a missing input, reorder by data flow, swap a step, remove a repeat), the proposed steps with all but swaps applied, and open_in_builder, a link that opens the proposal in the site's builder. Also returns an unsaved portable draft, selected skill inventory, review hash and handoff_check. Send full steps with pinned revisions instead of slugs for custom notes and dependencies. Suggestions and warnings never block the draft; show them and apply only what the user approves. Does not compose with our AI, save, execute or certify approval; an approval hash never authorizes external actions.

The roster a team job would get: one lead who owns the deliverable and the helpers it may call, matched from published roles and personas. start_job returns this for a team job. Review any unmapped catalogue gaps before accepting.

Without a reading: the read-back schedule for every open experiment and a routine prompt for a host scheduler (Claude Code scheduled task or cron), set up only if the owner agrees. With a reading: compares it with the registered baseline, minimum volume and guardrails; flags a guardrail breach at once; after the window, returns an honest verdict and the close payload for record_outcome_experiment. Budget and pause ideas come back as proposals for the owner; nothing is ever changed.

Record what the owner chose, the options considered (best-ranked first), what they passed over and why. Passing over the top option needs a reason. Returns a line for decisions.jsonl and, if a durable rule was stated, a learning candidate with where to save it.

Register a test before it goes live (action register: change, metric, baseline, window, minimum volume, guardrails), then close it (action close: observed, volume, guardrail results, verdict). Low volume forces INCONCLUSIVE; a broken guardrail is never a WIN. Returns a line for experiments.jsonl, a learning candidate and an anonymised outcome report the owner may choose to share.

Search AI Blocks in the shape ChatGPT deep research and company knowledge use: results with id, title and url. Read one with fetch. search_library does the same with filters and one-line summaries.

Look up one item (to start a piece of work, use prepare_job_context; for a business goal, start_outcome). Find skills, workflows, agents, prompts, minds, AI tools and mini tools by keywords and intent. Filters: type, job, level, tested, since (what's new). With no query, a job or since browses. Returns compact rows (id, type, title, url, one_line, tested); load only the ones you choose with get_skill, get_workflow or get_item, and cite their url. Plain search; does not use a server-side model.

Anonymous calls only prepare local preference files; they save nothing on the server. Signed-in connections require the selected workspace write consent. An ordinary one-off request never changes saved defaults.

Decides who does a job and says why. One kind of work: one lead, with the lead to load, the questions and the rubric. Mixed kinds of work, a large batch or mode team: one lead with helpers, the short team method and a state for advance_job. All work runs in the user’s AI; no server model calls or action authority.

For a goal or problem with no deliverable named yet, in the owner's own words ("more customers", "my ad costs are rising"). For a named piece of work, such as a page, ads, emails, an audit or a plan, use prepare_job_context. Returns a new engagement waiting for its brief, the next step and the files for outcomes/<id>/, with METHOD.md (brief frame with examples, research protocol, composition and run rules). include_example adds a worked example of a good brief, research and plan. No work runs here.