| name | reflect-and-improve |
| description | The reflection agent. Use it right after another skill produces an artifact such as a rewrite from polish-prose, a rebuttal, an abstract, a preflight fix, or a tailored draft to critique that output against the stated goal, the paper's .paper-memory/profile.yml, the relevant rubric, and past .paper-memory/lessons.md, then decide whether it actually improved. Trigger words include reflect, self-critique, did this get better, did my edit help, regression check, review my own output, iterate, refine, second pass, sanity-check the rewrite, and "is the new version actually better". Runs a bounded self-refine / Reflexion loop, guards against regressions, and appends durable, deduplicated, dated lessons to .paper-memory/lessons.md via a stdlib script so the toolkit gets sharper each paper. Reflection adds cost; it only pays off when there is a measurable target. |
Reflect and Improve
A meta-skill that runs after another skill in this toolkit produces an
artifact. It asks one question honestly: did the change actually make the
paper better, measured against the goal — or did it just make it different?
It critiques the new artifact against (a) the goal, (b) the paper's
.paper-memory/profile.yml, (c) the rubric that fits the task, and (d) past
.paper-memory/lessons.md; iterates within a hard budget; refuses to accept a
change that lowers a measurable score; and records what it learned so the next
paper starts smarter.
This is a copilot, not a pilot. It proposes; the author decides and edits.
When to use
- Just after
polish-prose, write-rebuttal, write-abstract,
preflight-check, tailor-to-venue, draft-related-work, or any skill that
emits an artifact, and you want a guarded second pass — "is the new version
actually better, or did I introduce a regression?"
- "Reflect on / critique / self-review this rewrite before I keep it."
- "Did my edit help?" / "did this get better?" / "should I revert?"
- To capture a durable lesson ("we keep over-hedging the abstract") so the
whole toolkit stops repeating the mistake on this and future papers.
When NOT to use (be honest about cost)
Reflection is an extra LLM pass plus author attention. It only pays off when
there is a measurable or rubric-anchored target. Skip it (or keep it to a
single quick pass) when:
- There is no goal to measure against — pure preference edits ("make it sound
nicer" with no rubric) reduce to taste; one read is enough.
- The producing skill already emits a hard score that did not change (e.g. a
preflight that was clean before and after). Re-running its checker IS the
reflection; do that instead of a narrative critique.
- The artifact is tiny (a one-line fix). The loop's overhead exceeds its value.
Say this to the user when they ask to reflect on something unmeasurable, then
offer the single-pass version.
Inputs
- The artifact to critique: the new text/file (and, ideally, the version
it replaced, so "better" can be a diff, not a vibe).
- The goal: what the producing skill was asked to achieve, in one
sentence ("cut the abstract to 200 words without dropping the contribution",
"answer R2's soundness concern").
.paper-memory/profile.yml — author/paper positioning. If it does not
exist, create it first with paper-profile (or hand-write the minimal
fields below). Reflection without positioning gives generic advice.
- A measurable target, where one exists:
- a script score (preflight error count,
polish-prose readability/hedge
counts, abstract word count vs venue bound, citation-verify failures);
- or a rubric in
references/ (or simulate-reviewers/references/rubrics.md)
scored 1–5 before and after.
.paper-memory/lessons.md — read at start so the critique personalizes
and does not repeat advice already logged.
The .paper-memory/ convention
A per-project directory in the user's paper working directory (never in
this repo). The skill creates it at use time. Tell the user to add
.paper-memory/ to their .gitignore unless they want positioning and
lessons versioned with the paper. It is local; nothing is uploaded.
profile.yml — vertical (systems/theory/applied/empirical/survey/position),
contribution_type, target audience & venue tier, risk appetite, writing
preferences (link to match-style), constraints. Minimal starter in
references/paper-memory-schema.md.
lessons.md — accumulated, deduplicated, dated lessons. Other skills
(preflight-check, polish-prose, verify-citations, simulate-reviewers)
APPEND here when they catch something and READ here at start. Format and
hygiene rules in the schema reference; enforced by scripts/reflect_log.py.
decisions.md — venue/track/positioning decisions with rationale (written by
select-venue / plan-submission; read here for context).
Process
The loop is Reflexion / self-refine, bounded: critique → revise → re-score
→ keep only if it measurably improved. Cap iterations; stop on no gain.
-
Establish the baseline — measure first. Before critiquing, capture the
current score of the artifact against its target. Pick the cheapest valid
measure: re-run the producing skill's checker if it has one (preferred —
deterministic), else score the relevant rubric 1–5. Record it:
python3 scripts/reflect_log.py score --memory .paper-memory \
--skill polish-prose --metric hedge-count --before 11
If there is no measurable target, say so, do one qualitative pass, and
stop — do not pretend a number exists.
-
Read memory. Surface recurring lessons so you neither repeat past advice
nor re-introduce a known regression:
python3 scripts/reflect_log.py recurring --memory .paper-memory
Read profile.yml. A "tighten every sentence" critique is wrong for a
position paper that values voice; a "add more hedging" critique is wrong for
a high-risk-appetite systems paper. Tailor to the profile.
-
Critique against the four anchors, concretely and per-finding:
- Goal — did it do the specific thing asked, fully and only that?
- Profile — does it fit the vertical, audience, venue tier, voice?
- Rubric — score the fitting dimensions (soundness, clarity,
novelty-framing, responsiveness for rebuttals) 1–5 vs the prior version.
- Lessons — does it repeat a logged mistake, or fix one?
Write each as {what regressed/improved, evidence, minimal fix}.
-
Decide — the regression guard is non-negotiable. Compute the after
score the same way as the baseline and compare:
python3 scripts/reflect_log.py score --memory .paper-memory \
--skill polish-prose --metric hedge-count --before 11 --after 6
The script prints the delta and a verdict. Never accept a change that
lowers a measurable score (more preflight errors, worse rubric mean, over
the word bound, new citation failures) even if the prose "feels" better — a
measurable regression beats a feeling. Outcomes: KEEP (measurably better) /
REVERT (measurably worse — recommend the prior version) / TIE (no measurable
change — defer to author taste, do not burn another iteration).
-
Iterate within budget. If KEEP but still short of target and under the
iteration cap (default 3 total passes), produce a focused revision
addressing the top finding only, then return to step 4. Stop when: the
target is met, the cap is hit, or two consecutive passes show no measurable
gain (diminishing returns — the honest signal to stop).
-
Append durable lessons. When the loop surfaces a pattern worth carrying
forward, log it — deduplicated and dated:
python3 scripts/reflect_log.py append --memory .paper-memory \
--skill polish-prose --scope recurring \
--issue "abstract over-hedges the contribution" \
--rec "state the result as a claim once, then qualify in the body"
--scope this-paper for paper-specific notes, recurring for cross-paper
patterns (these are never pruned). The script dedupes on
(skill, issue) and stamps the date.
-
Hygiene. Periodically prune stale this-paper lessons and enforce the
cap (recurring lessons are always kept):
python3 scripts/reflect_log.py prune --memory .paper-memory \
--keep 50 --stale-days 365
See references/reflection-loop.md for the
patterns this implements and their failure modes (reward hacking, sycophantic
self-approval, oscillation), and
references/paper-memory-schema.md for
the file formats.
Output
- A critique: per-anchor findings (goal / profile / rubric / lessons), each
with evidence and a minimal fix.
- A KEEP / REVERT / TIE verdict backed by a before→after number, not a
vibe. On REVERT, recommend the prior version and explain which measure dropped.
- Zero or more lessons appended to
.paper-memory/lessons.md.
- An honest statement of confidence and of what reflection could not measure.
Adapt to your discipline
The loop and memory are field-agnostic; only the rubrics and metrics are
CS-flavored. Swap in your field's review form for the rubric, your venue's
length/format limits for the script metrics, and your citation style for the
verify hook. The .paper-memory/ schema does not change.
Guardrails
- Never accept a measurable regression. A lower score is a REVERT, however
good the new version feels. This is the skill's whole point.
- Reflection is not free. State the cost; skip the loop when there is no
measurable target and offer a single pass instead — do not manufacture a
metric to justify iterating.
- Bounded loop. Hard iteration cap; stop on no measurable gain. Never spin.
- No self-sycophancy. Do not approve your own previous output by default;
the baseline measure, not your prior reasoning, is the judge.
- Never fabricate scores, rubric numbers, or "improvements" — measure or say
you could not. Route citation problems to
verify-citations, desk-reject
format problems to preflight-check.
- Never submit anything anywhere; never promise acceptance.
.paper-memory/
stays local — remind the user to .gitignore it.