| name | rl-review |
| description | Observe what the user actually did — from whatever evidence they provide (a report in their words, a file or export they point to, a screenshot, a paste, a link) plus the AI's own work records — and judge it against a life-harness goal's action plan and success criteria and the user's philosophy — honestly, not as a cheerleader — then log what happened into the action file and refine. Use when the user runs /rl-review, reports what they did / achieved, points at activity data to assess, or wants an honest check on progress over a time window. Reads goals/, actions/, philosophy/, and any AI work records under work/; the activity evidence comes from the source the user names; writes to actions/ and goals/. |
| metadata | {"loop":"life-harness","step":"4-review"} |
rl-review — observe the evidence, then tell the truth
This is the observe → reason turn of the loop. The user did things — on the
computer, in the world, or both — and this skill holds them against the goal's
timed steps, its success criteria, and their philosophy, then logs an honest
verdict into the action file. It is explicitly not a cheerleader: take the
user's account however they show it, hold it against their own plan, goals, and
philosophy, and tell them where they diverge — kindly, specifically, no spin.
How the evidence arrives is the user's choice, not a fixed pipeline: they can
tell you, or point you at an export, a screenshot, a log, a file, a paste, or a
link. Take whatever they provide and measure it against the plan.
Input
Whatever the user typed after /rl-review is a seed, not a strict argument: a
goal and/or window, a report of what happened, a pointer to evidence, or
nothing → ask which goal, over what window, and how they'll show what they did.
Steps
-
Locate the workspace (see rl-init conventions); if none is found, tell the
user to run /rl-init and stop.
-
Load the standard to judge against. Read the target goal under goals/open/,
its action file actions/open/<slug>.md, and all of philosophy/. Pull the
goal's success criteria and the plan's timed steps verbatim, and note
from "How progress is observed" how each was to be evidenced. Scan work/open/
and work/closed/ for any task.md whose goal:/action: points here: steps
the user delegated to the AI are evidenced by that record, not the user's report
— read it, and don't credit the user with the AI's work. The last ## Log
entry's timestamp is the default start of this review's window.
-
Fix the time window. Confirm start and end to the minute where it matters,
each with offset + IANA zone. If the user said "this week", state the concrete
window; if they changed timezones mid-window, note it.
-
Take in the evidence. Read (or ask for) whichever source the user is giving
— a direct report, a file/export on disk, or a paste/link (fetch it if you can;
if not, say so and ask for the content). Combine sources if offered. Pull out
concrete, checkable facts, not a vibe: what got done, how much / how often /
when, the measurable outcome, what got in the way. Push gently past "it went
okay" to specifics, but take a report at face value — you're not interrogating —
and record what was actually claimed, noting where the user is unsure rather
than rounding up. If no usable evidence is offered for a step, don't fabricate
it — ask for a source or record the step unverified. Sensitive data: exports
and screenshots may hold passwords or private messages — use them only to assess
this goal, never surface secrets in the log, and summarise ("~2h in the editor
on the project repo") rather than quoting raw captures.
-
Assess objectively, step and criterion by criterion. For each, judge done
/ partial / not done (or met / partial / not met) and cite the specific
evidence — user-did steps from the user's evidence, AI-did steps from the
work/ record. Hold steps to their time constraint (a "Mon–Fri, 2h before noon"
step done twice, late, is partial — say why). Distinguish what the evidence
shows from what it can't; note gaps instead of filling them with optimism.
-
Check philosophy, not just the number. Did the way the user acted honour
their values? Moving toward a goal by violating a principle is worth naming; so
is honouring the philosophy while missing the number.
-
Propose refinements when the evidence calls for it. Suggest concrete
changes — sharpen a vague criterion, adjust an unrealistic/too-easy target,
re-plan a step that isn't working, split or retire a goal, flag a
philosophy/goal tension. Recommend with reasons; the user decides. Route
criterion changes to /rl-goal, step changes to /rl-plan. If all criteria are
met, say so and propose marking the goal accomplished.
-
Log the verdict into the action file. No separate review file — append one
entry to ## Log of actions/open/<slug>.md, tick/adjust the affected Plan
checkboxes, and bump updated. The entry is delta-only (what changed since
the last entry — the criteria and plan already hold current truth) and
budgeted (a heading line + ≤5 one-sentence bullets; no paragraphs or
numbered sub-lists; one load-bearing "Philosophy check" line). Timestamps per
the workspace convention. Then keep the last ~3 entries in ## Log; move
older ones verbatim to actions/history/<slug>.md (append-only, not
auto-loaded) with a one-line pointer. Shape:
YYYY-MM-DD HH:MM ±HH:MM (Zone/Name) — [review] — window <start> to <end> — source: <how evidence was provided>
- Step / criterion — done | partial | not done — <specific evidence, one line>
- Philosophy check: <did the way they acted honour their values?>
- Verdict & refinements: <honest bottom line; any proposed goal/action change>
-
Update the goal, and close if reached. Append a one-sentence dated line to
the goal's ## Notes / history (rolling older entries to
goals/history/<slug>.md), and rewrite — don't append an aside to — any
criterion this review changes. If all criteria are met and the user agrees, set
status: accomplished and run the full close sweep (rl-init → Closing a
goal), not just the goal and action files: flip the action file (and any open
work/ task serving this goal) to done, move goal + action + those work
folders to closed/, and repoint every cross-reference (actions:, goal:,
work goal:/action:) from open/ to closed/. Then hand back to /rl-goal
or /rl-plan if refinements were accepted, or return to acting.