dimagi-internal/ace / skills/partner-on-connect-program
partner-on-connect-program
UnverifiedWork with a program team on a live Connect program as a standing project: answer questions, keep its decisions, and change what is live (apps, opportunities, bots, solicitations, dashboards). Use for program help outside /ace:run.
Partner on a Connect program
ACE is the program teams' AI partner for all things Connect, not only the builder of new programs (Jon, 2026-10-07, when the Kangaroo Care team asked to work with ACE on Labs metrics tooling). That work is ongoing, comes from several people, and outlives any single build. So it is a canopy project, with ACE's knowledge of Connect layered on top.
This skill describes the extension. The project and task mechanics are fleet-wide: follow
skills/task-tracker, which reads canopy's agent-core/task-tracker.md, and
skills/gdoc-writer for deliverables. Do not re-implement either one here.
A program project is not an opp
| Program project (this skill) | ACE opp (/ace:run) | |
|---|---|---|
| Subject | A live Connect program and its team | One new opportunity built from scratch |
| Lifetime | As long as the team works with ACE | One run, then frozen |
| State | canopy-web project + tasks; Drive Projects/<name>/ | ACE/<opp>/runs/<run>/run_state.yaml |
| Human surface | The board, the threads, the deliverables | The run's ace-web page |
| Source context | The program context doc (below), read live | The frozen inputs/ manifest |
Bridge, don't merge. When program work needs a new build (a new opportunity, or a fresh app
for a new partner), start a normal ACE opp: create ACE/<opp>/inputs/ holding shortcuts to
the program's living docs, and run it through /ace:run. Then link that run's ace-web page to
the project. Program work never writes a run_state.yaml, and an opp never reads the project's
board. When asked about a run, read it through its ace-web page and artifacts. Don't
reconstruct it.
The project
- One project per program, named as the team names the program, for example
Kangaroo Care. The board project and the Drive folder share that one name (task-tracker § Projects). If a folder already exists under an old name, rename it to match rather than starting a second folder. - One task per thread of work (a tool being built, a question being chased), filed into the project. Owner is the program-team person who owns the outcome.
- Outcome on the project says what the partnership is for right now, in the team's words.
The program context doc
One Google Doc in the project folder, titled <Program> — program context. It holds only
pointers and decisions. Facts that live in another system go in as a link or an ID, never as
a copy. The team edits it. ACE appends decisions with the date and who made them, and never
rewrites what someone else wrote.
| Section | What goes in it | Preferred form (never stale) |
|---|---|---|
| Names | What the team calls the program and its parts, and any retired names | Plain text. Retired names are still searched for, never written in prose |
| Scope | Connect org(s) and program, opportunities in and out of scope, with the reason | Connect org slug + opportunity IDs |
| Apps | Learn/Deliver apps per opportunity, and which build counts | HQ domain + app ID, and "latest released" unless stated |
| AI coaches and bots | Every OCS bot the program uses (coaches, Q&A bots), what each is for, and which Labs reports trigger it | OCS team + bot id/public id per bot. There is rarely just one, so a question about "the coach" is resolved here first |
| Metrics | Indicator definitions, reports, cohorts | Labs registry, workflow and cohort IDs. A spreadsheet is the drafting space; the registry is the authority once adopted |
| Living docs | Design docs, protocols, the team's working sheets | Drive links shared with ace@dimagi-ai.com (commenter) |
| Decisions | What was decided, by whom, when | Dated lines. A decision made only in Slack or email reaches no future session unless it lands here |
| Data access | What ACE may read, and who approved it | One line per grant, with the owner's approval |
| People | Who is on the team and what each person owns | Names and emails |
| Builds | ACE opps and runs started for this program | ace-web run-page links |
Every session
-
Find the project.
canopy agent projects --slug ace --active, then read its tasks and the context doc. If the request doesn't map to a project, ask which program it is. Don't guess from an acronym. -
Resolve which instance the question is about. "The coach", "the app" or "the report" means one of the program's instances in the context doc. Name it before answering, or ask. Never answer for ACE's generic template as if it were the program's.
-
Re-read live only what the question needs. Use the context doc's pointers:
- the HQ app via the
commcare_*atoms - the opportunity via the
connect_*atoms - metrics via the
connect-labsregistry and workflow atoms - docs via Drive
Search under the program's retired names too.
- the HQ app via the
-
Say what you read and when in the answer. A dated snapshot from the folder is reference, never the source. The 2026-08-12 KC field audit, for example, describes the apps as they were that day.
-
Write back. Log the turn on the task (
canopy agent turn … --task T<N>). Put deliverables on the project's links. Append any decision to the context doc.
Changing a live program
Most requests a program team makes are changes to something already live: "add a question to
the Deliver form", "lower the pass mark", "the coach should mention X", "add a payment unit for
the follow-up visit". These are modifications to existing objects. They are not new builds.
Never route one into /ace:run, which creates a fresh opportunity and a fresh app and would
leave the partner with two of each.
What counts as a change: any write to an object listed in the context doc's Scope, Apps, AI coaches and bots, or Metrics. That covers an app edit plus re-release, an opportunity's verification flags, payment units, dates or budget, the Learn passing score, a bot's prompt, knowledge or tools, a solicitation's fields, and a Labs workflow's opportunities or template. Reading and answering are not changes.
Whose account makes it. A labs change (a workflow, a dashboard, the semantic layer) is made as ACE unless the person said "as me". Either way, name the account before writing: skills/_acting-as-you.md.
Hiding a run is not closing an opportunity. Connect has no archive or delete for an opportunity
(deactivating is the only move, connect_update_opportunity active:false, and the opportunity
stays listed). To tidy what people SEE, archive the run in ace-web: POST / DELETE /api/w/{ws}/opps/{opp}/runs/{run}/archive with ACE's ACE_WEB_PAT_TOKEN, listed at
/w/{ws}/archive. It writes only an ace-web row, so Drive, Connect, HQ and OCS are untouched and
it needs none of the checks below. Say plainly which of the two was done; Connect still lists
the opportunity.
Procedure:
-
Name the object and its current state. Resolve it from the context doc (§ Every session step 2), then read it live and record the before state: the field values, the app build number, the bot version.
-
Check the platform's limits before proposing anything. The playbook index (
playbook/README.md, "how do I do X on surface Y") and the matchingplaybook/integrations/<surface>.mdsay which operations exist and which do not. The traps that make a plausible change wrong are:- A
DeliverUnitis shared across opportunities wired to the same released Deliver app (ace#573). A payment unit cannot be bound to a deliver unit another opportunity already holds, and no atom deletes a payment unit. Adding or changing payment units on a live opportunity is constrained, so check before promising it. - The Learn
passing_scorelives on the sharedCommCareApprow (ace#1350). Changing it withconnect_set_learn_passing_scorechanges the gate for every opportunity in the org wired to that Learn app. Name those opportunities in the proposal. - An opportunity's holding org is create-time only.
target_organization_slugcannot be changed (connect_update_opportunityhas no org field). A wrong org means a new opportunity, which is a build, not a change. - An app edit only reaches workers once it is released. The path is a Nova edit
(
nova:edit, or the private-work atoms), upload to the same HQ app ID, re-apply the HQ settings, rebuild and release, then confirm the opportunity sees the new deliver units. Whether a live opportunity picks up a re-release on its own is not documented. Read it back rather than assuming it. - An OCS bot's public link serves only the published version. Republish after any
prompt or knowledge change (
ocs-knowledge-refreshcovers knowledge).
- A
-
Propose the change with before and after. One message to the requester or owner gives the object, its before state, the intended after state, what else it touches (other opportunities on the same app, workers mid-Learn, live Labs reports), and how it will be verified. The owner approves any outbound or shared act (manual-mode turns), and a change to a live opportunity or a released app is one. Program-team members can direct the work, but approval follows § Guardrails, Who steers.
-
Make the change through the modify skill for that object. Each one takes explicit IDs (from the context doc), refuses what the platform cannot do, and reads back:
Object Skill An app's forms, fields, logic or case config, shipped to its live opportunity app-modifyAn opportunity's verification flags, payment units, dates, budget, Learn passing score connect-opp-modifyA new partner org's own opportunity on the program (invite, accept, fresh apps, opp, payment units) connect-partner-opp-createA bot's prompt, description, tools or pipeline ocs-bot-updateA bot's knowledge ocs-knowledge-refreshA solicitation's deadline, text, questions, criteria, visibility or status solicitation-editA labs report's opportunity list, or freezing / refollowing its template labs-workflow-repointCase sharing between workers app-case-sharingDon't hand-roll a procedure one of these owns. These skills cannot ask the human (a dispatched copy has no
AskUserQuestion), so hand each one the approval from step 3 in the dispatch. Where a skill still reads its IDs only from a run's artifacts, pass the IDs from the context doc and note the gap on the task. -
Verify against a fresh read. Never trust the write's status alone: re-read the object and compare it with the agreed after state.
-
Log it. Append a dated line to the context doc's Decisions: what changed, the before and after values, who asked, who approved, and the evidence (a link or ID). Log the turn on the board task. A change that lands only in an email or a session reaches no future session.
A change too large to finish in one turn (an app edit plus release, a multi-step
reconfiguration) is dispatched to its own session with the project, the context-doc link and
the asks verbatim, the same way inbox-triage dispatches a run.
When a build becomes a program project
A program ACE built through /ace:run becomes a program project once it is live, meaning
after Phase 9 go-live (llo-launch) has activated the partner's opportunity. From then on,
requests about it are program work under this skill, not steps of the run:
- Open the project (§ The project) if none exists, and seed its context doc's Scope, Apps and
AI coaches and bots from the run's
run_state.yamlproducts. Copy the IDs as pointers, not the artifacts. - Put the run's ace-web page (
ace_web_summary_url) in the context doc's Builds section, so the build history stays one click away. - The run itself stays frozen. A change to the live program is never written back into that
run's
run_state.yaml. It is logged in the context doc's Decisions.
Before go-live the run still owns its objects, and changes go through the run (its pause points
or /ace:step).
Guardrails
- Data. People may share with ACE only what they would share with their own AI assistant
(owner, 2026-10-07). ACE runs in an ordinary Claude session with no extra data controls, and
ZDR mode is reserved for Jon. Point people to that rule. Don't make up a stricter or looser
one. What ACE's own accounts can reach is a separate question, so check it live, not from
memory: on 2026-10-07 its Labs account could list the real KC programs through the Dimagi
org, after an August write-up said it saw only synthetic copies. Granting ACE an account
membership, for example adding
ace@dimagi-ai.comto a program's Connect org, is the owner's decision. Record it in the context doc's Data access section. - Who steers. Program-team members can direct the project's work. Outbound and shared acts
need the owner's approval in a manual-mode turn: publishing or cloning in Labs, sharing a
doc, emailing a partner, granting access. This follows
skills/explain-ace§ 2 and canopy-web'sdocs/architecture/access.md. - Documents shared with ACE. Ask the person to share with
ace@dimagi-ai.comas commenter. A Drive "file not found" for both of ACE's identities means the file hasn't been shared with ACE. It does not mean the file is gone. - No inferred backstory. A target, threshold or scope rule that isn't in the context doc or a linked source is a question for the team, not a default.
- Terminology. Use the program's current names in everything a human reads. Identifiers (folder names, app files, slugs) stay as they are.
Related skills
skills/task-tracker: board and project mechanics.skills/gdoc-writer: publishing deliverables intoProjects/<name>/.skills/explain-ace: interaction and security model questions.skills/inbox-triage: how inbound mail is routed to a program project (§ 2b) and handled as Program work (§ 2d).playbook/README.mdandplaybook/integrations/*.md: what each platform can and cannot do when changing a live object./ace:run,/ace:program-update: when program work turns into a new build.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 192k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 192k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 192k | Updated | Claude CodeCodex | Verified | Install |