Make what works part of how your team works.

A correction that helped one session should help the next. Team playbooks give a useful lesson an owner, a trigger, a scope and a version, so reuse is a decision the team makes rather than something that happens by accident.

Team playbooks · Repository context Sample

One lesson, from capture to a considered decision.

The same repository is rediscovered at the start of a run Seen in 62 eligible sessions · mostly Faster CI/CD · August 2026
Source sessions attached to the lessonUsage records · airis · Engineering
Recorded
Owner namedFrank · Platform engineer
Assigned
Draft v1 filed in the team librarySkill · Applied to no runner
Draft

A lesson keeps its source sessions, its owner and its version · Sample workspace, synthetic data

Current practice Rediscover the repository in every run 62 eligible sessions · Usage estimate
Prepared change Reuse verified repository context Skill · Draft v1 · Future owned runs
Acceptance criteria unchanged The same recorded checks and human review as current practice

Instructions are reviewed in the workspace · The preview carries only what the team needs to decide

  1. 1 Understand
  2. 2 Prepare
  3. 3 Compare
  4. 4 Apply & learn

Repository context

Compared, not applied
Acceptance pack attachedSame criteria as current practice
Ready
12 paired CI tasks selected4 held-out cases · Unfinished tasks retained
Ready
Compared on 12 paired CI tasksResult with Frank for a decision
Compared

Nothing is applied by being prepared · A person applies a reviewed version to future owned runs · Concept data

01 / 05 The lesson

A useful lesson should outlive the session that produced it.

  • 01 Every playbook has an owner, a trigger, a scope, a version and an explicit application state.
  • 02 A prepared version is compared with current practice on the same tasks before anyone relies on it.
  • 03 Reading a recommendation changes nothing. A runner changes only when a person applies a reviewed version.

02 / 05 Capture, prepare, compare

Three steps from a lesson to a shared practice.

A lesson can take the form of a skill, a workflow or a hook. What the team needs to know is the same in each case: where it came from, who owns it, what it applies to and whether it has been compared.

01 Capture

Keep the lesson, and where it came from.

When a correction in one session would help the next, agentacct captures it with its source sessions, names an owner and files it in the team library as a draft. A lesson in the library is visible to the team. It is not applied to anything by being there.

  • Source sessions stay attached to the lesson
  • One named owner per playbook
  • Drafts are visible before they are used
Cost intelligence
Library · 6 lessons · 3 shown Sample
Skill · Draft v1 Compared, not applied

Repository context

Owner Frank · Scope Future owned runs · Trigger Task start in a mapped repository
62eligible sessions
Workflow · v2 Compared, awaiting decision

Focused validation

Owner Alex · Scope Runs that edit tested code · Trigger After each edit
68eligible sessions
Hook · v1 Prepared, not applied

Environment preflight

Owner Sam · Scope Future owned runs · Trigger Before the first command
62eligible sessions

Eligible sessions counted from August usage records · Three more lessons not shown · Sample

02 Prepare

Scope the change before anyone uses it.

A prepared playbook states its trigger, its scope and its owner, and carries the acceptance criteria that current practice already has to meet. The generated instructions stay in the workspace, where the owner reviews them. The preview shows only what the team needs to decide.

  • Trigger, scope, owner and version on every draft
  • Acceptance criteria carried over unchanged
  • Instructions reviewed in the workspace, not in a preview
Delivery speed
Playbook preview · Repository context Sample

Repository context

Draft v1
Trigger Task start in a repository the team has already mapped
Scope Future owned runs · airis · Engineering
Owner Frank · Platform engineer
Version Draft v1 · First version, replaces nothing
Source 62 eligible sessions · August 2026
Acceptance criteria unchanged Recorded checks and human review stay exactly as they are for current practice

Trigger, scope, owner and version are the preview · Full instructions never leave the workspace · Sample

03 Compare

Compare before use. Apply on purpose.

A version runs against the same paired tasks, revision, environment and acceptance criteria as current practice. Unfinished tasks and held-out cases stay in the result. Only after review does a person apply the version to future owned runs, and the library records that state beside the version.

  • Same tasks, same criteria, unfinished work retained
  • Application state recorded per version
  • Applied means future runs, never past evidence
Execution quality
Compare · Repository context · Draft v1 · 12 paired CI tasks Concept data
Frank · Owner · Skill · Draft v1 Compared, awaiting decision
Tasks accepted
7 / 1210 / 12
Time to decision (unfinished at cap)
2h 29m1h 44m
Cost per accepted task
$48.12$26.96

Same input, revision, environment and acceptance criteria as current practice · Concept data, not a measured result

03 / 05 The library

Six lessons in the library. Three of them, up close.

Each card shows what the team needs to decide about a playbook and nothing it does not: owner, scope, trigger, version and application state. Sample

01 · Skill · Draft v1 Repository context

Reuse verified repository context at the start of an owned run instead of rediscovering it.

Compared, not applied
Owner · Frank Scope · Future owned runs Trigger · Task start in a mapped repository
02 · Workflow · v2 Focused validation

Run the checks an edit touches between edits and keep the full suite for the final pass.

Compared, awaiting decision
Owner · Alex Scope · Runs that edit tested code Trigger · After each edit
03 · Hook · v1 Environment preflight

Confirm the environment before the first command, so agents stop retrying setup errors.

Prepared, not applied
Owner · Sam Scope · Future owned runs Trigger · Before the first command

04 / 05 What a playbook will not do

Proposals stay proposals until a person applies one.

The vocabulary is strict because the decisions are real: which runs a version touches, on what evidence, and who decided.

No silent activation

Reading changes nothing.

Opening a recommendation, a draft or a comparison never modifies a runner. Application is a separate, recorded action by a person.

No unversioned reuse

Every version is named.

A change to a playbook is a new version with its own comparison. Runs record which version they used, so evidence stays attributable.

No cherry-picked wins

Comparisons keep the failures.

Paired tasks, retained unfinished work and held-out cases stay in the result. A better number on fewer tasks is not a result.

No instructions in the open

Detail stays with the owner.

Generated instructions and full workflow definitions live in the workspace, where the owner reviews them. A preview carries only what a team needs to decide.

See it with your own lessons in mind

Make the next run start
one lesson ahead.

A guided look at how a lesson becomes a compared, versioned playbook for a sample cohort, and a conversation about your teams and tools.