ax:extract-workflow - reconstruct the recipe behind a shipped artifact
Given a deliverable (a demo, a PR, a feature) the user wants to know:
what skills, in what order, produced it? This skill resolves an anchor
(date or sha), windows the relevant sessions, and narrates the ordered
skill arcs in plain text. Eats its own dogfood: it is itself a framing
skill.
Assumes
(axctl) is on PATH. If
fails, tell the user
to check
docs/development.md#setup
(DuckDB dylib setup - no daemon
required) and stop.
When to fire
ONLY on explicit reconstruction triggers:
- "what made <X> work" / "how did we ship <Y>"
- "extract workflow from <date>" / "extract workflow from <sha>"
- "what was the workflow around <topic>"
- "reconstruct the recipe for <artifact>"
- "show me how I built <feature>"
Do NOT fire on generic "what did I do today" or "show recent activity".
That is
territory, not this skill.
Step 1 - resolve the anchor
Decide one of three modes based on what the user gave you:
| User said | Mode | Action |
|---|
| commit sha (full or short) | sha | use it directly |
| date or date range (YYYY-MM-DD) | date | use ax sessions around <date>
|
| topic / feature / artifact name | topic | ax recall "<topic>" --sources=commit --json
|
| "this repo, recently" | pwd | ax sessions here --days=14
|
For topic mode pick the most relevant sha and proceed in sha mode. If
results are ambiguous, ask the user to pick one before continuing.
Step 2 - window the sessions
Pick the right command for the resolved anchor:
- sha mode:
ax sessions near <sha> --json
- date mode:
ax sessions around <date> --days=3 --json
- pwd mode:
ax sessions here --days=14 --json
Pick the N most relevant sessions from the JSON (default N=5). Bias to
sessions with the highest turn counts and that touch files related to
the artifact.
Step 3 - inspect each session
For each picked session:
ax sessions show <id> --json
ax sessions show <id> --by-role # optional, see Step 4
Read the
and
arrays. If a subagent's
work looks central to the artifact, drill in:
ax sessions show <id> --expand=<subagent-uuid>
Step 4 - narrate
Produce two artifacts inline (no files written - keep the answer in chat).
4a. Ordered skill arc
Lead with the framing skill that opened the work (use
output
if available - the
group goes first). Then execution skills,
then verification. For each skill, name it and write one line on what
it produced.
Example:
1. brainstorming -> defined the workflow extraction problem
2. writing-plans -> turned 13 grilled questions into a plan
3. subagent-driven -> executed the plan as 17 typed-CLI tasks
4. code-review -> two-stage spec + quality review per task
5. test-driven-dev -> 100% green throughout
4b. Key decisions
Pull 2-4 turn excerpts where the user steered the work. Use:
ax recall "<keyword>" --skill=<framing-skill> --limit=5 --json
Quote one line per decision; cite the session id.
4c. Reproducer brief (optional, if user asks)
A one-paragraph "to do this again, you would" summary - the skills, the
order, and the key user inputs at the steering points.
When to recommend ax skills classify
If
ax sessions show <id> --by-role
returns many skills in the
group, the role-weighted output above will be noisy.
Suggest the user run
once - it generates briefs
the user can fill in to seed roles. Don't block on it; the
artifact-reconstruction story still works without roles.
Output contract
Write everything inline. Do not create files under
or
anywhere else. Do not modify the repo. This is a read-only,
reconstruction-focused skill.
If the user asks for a permanent recipe, recommend they paste your
output into
themselves - ax intentionally has
no recipe save format yet (per the workflow-extraction-frictions plan).